# Almond Robotics > Almond Robotics builds Axol, a dual-arm robot for builders working on physical AI. Two 7-DOF arms with 860mm reach and 6.5kg peak payload each, a 500Hz control rate, and an open-source Python SDK with LeRobot bindings and WebXR VR teleoperation. Designed and assembled in San Francisco, California. Contact: hello@almond.bot ## Products ### Axol (from $9,499) A dual-arm robot built for physical AI. Two 7-DOF arms with 860mm reach and 6.5kg peak payload each, deployment-ready out of the box with an open-source SDK. URL: https://almond.bot/axol - Configuration: Dual-arm robot - Arm Reach: 860mm - Degrees of Freedom: 7 per arm - Peak Payload: 6.5kg per arm - Nominal Payload: 4kg per arm - Repeatability: 1mm - Weight: 19.5kg - Power Input: 24V, 60A peak power - Control Bus: 500Hz CAN - Camera Interface: 2× internal wrist FAKRA GMSL 2.0 - I/O: USB-C (CAN control), XT90 (power) - Gripper: 2-finger gripper, replaceable tips or full swap - Mounting Holes: Base, wrist cameras, head camera - Construction: Steel, Aluminum, TPU - Color: Powder-coated matte black - Origin: Designed & assembled in San Francisco ### Jelly Mobile (from $7,999) Jelly is a mobile base for Axol with holonomic movement, integrated batteries, and a telescoping lift. Jelly takes Axol out of the work cell. Drive it anywhere in your facility, run all day on battery, and use the lift to pick something up off the floor and place it on the top shelf. Designed for Axol, but the mounting plate works with any robot. URL: https://almond.bot/jelly-mobile - Type: Powered mobile base with lift - Compatibility: Best with Axol, works with any robot - Footprint: 600 × 600mm - Battery: 2.4 kWh - Runtime: 4–10 hours - Lift Travel: 610mm (2ft) - Chest Height (min): 760mm (2.5ft) - Chest Height (max): 1,370mm (4.5ft) - Lift Speed: 40mm/s (1.57in/s) - Drive Speed: 0.7m/s - Lead Time: 1 week ### Ox Cart (from $999) The Ox Cart is a rugged, height-adjustable platform that gives Axol a stable footing and lets you reposition it across your workspace in seconds. Sit-to-stand height adjustment, four leveling casters, and a 60x60cm footprint that fits through standard doors make it ideal for flexible installations, while T-slot mounting keeps it easy to customize for your application. URL: https://almond.bot/ox-cart - Type: Height-adjustable mobile base - Compatibility: Axol dual-arm robot - Footprint: 60x60cm (fits standard doors) - Height Adjustment: Sit-to-stand range - Mobility: 4 leveling casters - Weight: 54.5kg - Mounting: T-slot extrusion - Mounting Holes: Base mounting pattern - Construction: Powder-coated steel - Color: Matte black - Origin: Designed & assembled in San Francisco ### Owl Mount (from $199) The Owl Mount is a rigid 400mm 8080 aluminum T-slot extrusion column that bolts Axol to a table, workbench, or fixture. Base brackets include clearance holes for 2020, 4040, and 4545 extrusion patterns, so it drops onto most T-slot tables without drilling. It's the lowest-cost way to give Axol a permanent home in your workspace. URL: https://almond.bot/owl-mount - Type: Static table/fixture mount - Compatibility: Axol dual-arm robot - Extrusion: 8080 aluminum T-slot - Height: 400mm - Base Footprint: 275mm wide - Fixture Holes: 2020 (26× M5), 4040 (12× M8), 4545 (12× M8) - Mounting: Table or fixture bolt-down - Construction: Aluminum extrusion - Origin: Designed & assembled in San Francisco ### Mantis UMI (from $1,799) Mantis UMI is a pair of handheld data collection grippers with the exact same grippers as Axol. Collect accurate bimanual manipulation data in the real world, 10x faster than teleop. Axol not required. Data transfers directly to Axol. URL: https://almond.bot/mantis-umi - Configuration: Handheld gripper pair - Sold As: Pair (2 grippers) - Gripper: 2-finger, Axol-matched geometry, swappable tips - Camera Mount: GMSL2 wrist-view mount for ZED X One S - Tracker Mount: Vive Tracker 3.0 / Quest 3 controller - Origin: Designed & assembled in San Francisco ### PSU (from $249) Power for Axol, sold on its own. Pick the Standard PSU, the same 1500W (24V, 62.5A) desktop supply included with every Axol, as a spare or replacement, or the Industrial PSU with a dedicated E-Stop, an integrated power switch, and a brake resistor for permanent installations and industrial cells. URL: https://almond.bot/psu - Type: Desktop power supply - Output: 24V, 62.5A (1500W) - Output Connector: XT90 - Compatibility: Axol dual-arm robot ### Camera Kit (from $1,499) The Camera Kit gives Axol its eyes: two ZED X One S Fisheye wrist cameras and one ZED X Mini stereo head camera, all global shutter, with Almond's head and wrist mounts and FAKRA GMSL2 cabling. Sold on its own for Axol owners adding vision later or replacing a camera. URL: https://almond.bot/camera-kit - Type: Camera kit - Wrist Cameras: 2× ZED X One S Fisheye (global shutter, 204° FOV) - Head Camera: 1× ZED X Mini (stereo, global shutter) - Interface: FAKRA GMSL2 - Mounts: Axol head and wrist mounts included - Compatibility: Axol; Jetson Orin NX 16GB, Jetson Orin AGX 64GB, Jetson Thor AGX 128GB ### Compute (from $1,999) Onboard compute for Axol and Jelly Mobile, sold on its own. Pick the NVIDIA Jetson that fits your workload: the Jetson Orin NX 16GB for camera capture, the Jetson Orin AGX 64GB for headroom with the full camera kit, or the Jetson Thor AGX 128GB for running VLA policies with every camera on one box. All work with the open-source Axol SDK. URL: https://almond.bot/compute - Module: NVIDIA Jetson Orin NX 16GB - AI Performance: 100 TOPS - Memory: 16GB - Storage: 256GB SSD - Camera Inputs: GMSL2 - Compatibility: Axol, Jelly Mobile, Mantis UMI # Robotics & Physical AI Glossary Definitions of robotics and physical AI terms, from https://almond.bot/glossary. ## Glossary: A ### A* Algorithm Category: Navigation & SLAM URL: https://almond.bot/glossary/a-star-algorithm A* is a best-first graph search algorithm that finds a least-cost path between nodes by expanding the frontier in order of estimated total cost, the sum of the cost-so-far and an admissible heuristic estimate of the remaining cost. Introduced by Hart, Nilsson, and Raphael in 1968 during work on the Shakey robot project, it is guaranteed to return an optimal path when the heuristic never overestimates. Variants such as weighted A*, D* Lite, and Hybrid A* extend it to replanning and kinodynamic settings. Why it matters for physical AI: Grid- and lattice-based A* variants remain the workhorse of global path planning in mobile robot autonomy stacks, providing predictable, optimal routes that learned local policies can then track and refine. ### A3C (Asynchronous Advantage Actor-Critic) Category: Robot Learning URL: https://almond.bot/glossary/a3c A3C is a deep reinforcement learning algorithm, introduced by Mnih et al. in 2016, that runs many actor-learner threads in parallel, each interacting with its own copy of the environment and asynchronously updating shared policy and value networks. The advantage estimate reduces gradient variance, while parallel exploration decorrelates training data without a replay buffer. It was largely superseded by synchronous variants such as A2C and later by PPO. Why it matters for physical AI: Parallel actor-learner architectures pioneered by A3C underpin modern large-scale simulation training for locomotion and manipulation, where thousands of environment instances feed a single policy update loop. ### ABB IRB Category: Hardware URL: https://almond.bot/glossary/abb-irb ABB IRB is the product family designation for industrial robot arms from ABB, one of the largest industrial robotics manufacturers. The line traces back to the IRB 6 of 1974, widely cited as the first all-electric, microprocessor-controlled industrial robot. Modern IRB models span payloads from a few kilograms to over half a tonne and are programmed via ABB's RAPID language and RobotStudio environment. Why it matters for physical AI: Installed fleets of arms like the IRB series represent an enormous deployment surface for learned policies, and their rigid, high-precision designs contrast with the compliant hardware often favored for contact-rich learned manipulation. ### Accelerometer Category: Hardware URL: https://almond.bot/glossary/accelerometer An accelerometer is a sensor that measures proper acceleration, the acceleration of a body relative to free fall, typically along one to three orthogonal axes. MEMS accelerometers sense the deflection of a microscopic proof mass and are cheap, small, and ubiquitous. In robotics they are almost always packaged with gyroscopes in an inertial measurement unit (IMU) and used for state estimation, vibration monitoring, and impact detection. Why it matters for physical AI: Accelerometer data feeds the high-rate proprioceptive state estimates that legged locomotion controllers and learned policies depend on, and it provides contact and slip cues that cameras cannot capture at millisecond timescales. ### Ackermann Steering Category: Locomotion URL: https://almond.bot/glossary/ackermann-steering Ackermann steering is a steering geometry for wheeled vehicles in which the front wheels are turned at slightly different angles so that their axes intersect at a common instantaneous center of rotation, allowing all wheels to roll without lateral slip. Vehicles using it are nonholonomic: they cannot translate sideways, which constrains feasible paths. The kinematic bicycle model is the standard simplification used for planning and control. Why it matters for physical AI: Autonomous vehicles, delivery robots, and outdoor platforms with car-like bases require planners and learned policies that respect Ackermann constraints, since paths feasible for differential-drive robots may be undrivable for them. ### ACT (Action Chunking with Transformers) Category: Robot Learning URL: https://almond.bot/glossary/act ACT is an imitation learning algorithm, introduced by Zhao et al. in 2023 alongside the ALOHA teleoperation system, that trains a transformer to predict chunks of future actions rather than single steps. It is structured as a conditional variational autoencoder over demonstrations, and at test time overlapping chunks are blended by temporal ensembling. ACT achieved fine bimanual manipulation from only tens of demonstrations on low-cost hardware. Why it matters for physical AI: ACT demonstrated that chunked action prediction mitigates compounding errors in behavior cloning, a design insight now standard in manipulation policies and vision-language-action models trained on teleoperated data. ### Action Chunking Category: Robot Learning URL: https://almond.bot/glossary/action-chunking Action chunking is the practice of having a policy predict a sequence of future actions in one inference step, which the robot then executes open-loop or with periodic replanning, instead of predicting one action per observation. Popularized by ACT and adopted by diffusion policies and most vision-language-action models, chunking reduces compounding distribution shift, better captures temporally correlated demonstrator behavior, and amortizes expensive inference over many control steps. Why it matters for physical AI: Chunking is what lets policies with slow, large backbones control robots at practical rates, and choosing the chunk length trades off reactivity against smoothness in deployed manipulation systems. ### Action Distillation Category: Robot Learning URL: https://almond.bot/glossary/action-distillation Action distillation is the transfer of action-selection behavior from a teacher policy into a student policy by training the student to match the teacher's output actions or action distributions. Common instances include distilling a privileged simulation policy with access to ground-truth state into a deployable policy using only onboard sensors, and compressing a large or slow model into a smaller, faster one. Why it matters for physical AI: Teacher-student distillation is the standard bridge from privileged simulation training to real-world sensing in legged locomotion, and it shrinks large foundation policies to sizes that meet onboard latency budgets. ### Action Expert Category: Foundation Models URL: https://almond.bot/glossary/action-expert An action expert is a dedicated subnetwork within a vision-language-action model that specializes in producing continuous robot actions, attending to the backbone's representations while keeping its own separate weights. The design was popularized by Physical Intelligence's pi-0 model, where a flow-matching action expert generates high-frequency continuous action chunks alongside a pretrained vision-language backbone, analogous to a mixture-of-experts partition. Why it matters for physical AI: Separating action generation from language-vision processing lets one component preserve web-scale semantic knowledge while the other models fine, high-rate control, a pattern now common across frontier robot foundation models. ### Action Head Category: Foundation Models URL: https://almond.bot/glossary/action-head An action head is the output module of a learned robot policy that maps backbone features to executable actions. Designs vary widely: discrete token heads that emit binned actions autoregressively (as in RT-2), regression heads producing continuous values, and generative heads based on diffusion or flow matching that sample from multimodal action distributions. The choice of head strongly affects inference speed, action fidelity, and the ability to represent multiple valid behaviors. Why it matters for physical AI: Head architecture is one of the most consequential design decisions in vision-language-action models, determining whether a policy can express multimodal strategies and run within real-time control budgets. ### Action Horizon Category: Robot Learning URL: https://almond.bot/glossary/action-horizon An action horizon is the number of future timesteps covered by a policy's predicted action sequence, or executed before replanning. Chunked policies often distinguish a prediction horizon (how many actions the model outputs, commonly around 16 to 100 control steps) from an execution horizon (how many are actually applied before the next inference call). Longer horizons yield smoother, more committed behavior; shorter ones react faster to disturbances. Why it matters for physical AI: Tuning the execution horizon is a core deployment knob: too long and the robot ignores perturbations mid-chunk, too short and inference latency dominates, so practitioners tune it per task and per platform. ### Action Space Category: Robot Learning URL: https://almond.bot/glossary/action-space An action space is the set of all commands a policy can issue to a robot, together with their parameterization. Common manipulation choices include joint positions, joint velocities, joint torques, and Cartesian end-effector deltas plus a gripper command; locomotion policies typically output joint position targets tracked by low-level PD controllers. The space may be continuous, discrete, or a hybrid, and its frame, dimensionality, and rate shape what behaviors are learnable. Why it matters for physical AI: Action space choice often matters as much as architecture: end-effector deltas transfer more easily across embodiments, while joint-space actions permit finer whole-arm control, directly affecting cross-robot generalization of foundation policies. ### Action Space Normalization Category: Robot Learning URL: https://almond.bot/glossary/action-space-normalization Action space normalization is the rescaling of robot action values into a standard numeric range, commonly minus one to one or zero mean and unit variance, using statistics computed over the training dataset. Normalization equalizes the loss contribution of dimensions with very different physical units, such as gripper width versus joint radians, and is required when actions are discretized into fixed bins for token-based policies. Statistics must be stored and inverted exactly at deployment. Why it matters for physical AI: Mismatched normalization statistics between training and inference are a notorious silent failure mode when fine-tuning or deploying pretrained policies, producing erratic motions that resemble hardware faults. ### Action Tokenization Category: Foundation Models URL: https://almond.bot/glossary/action-tokenization Action tokenization is the conversion of continuous robot actions into discrete tokens so that language-model architectures can predict them autoregressively. The simplest scheme bins each action dimension uniformly, as in RT-1 and RT-2, which mapped actions onto reserved text tokens. Later approaches compress action chunks more efficiently; the FAST tokenizer, for example, applies a discrete cosine transform followed by byte-pair encoding to represent high-frequency chunks with few tokens. Why it matters for physical AI: Tokenization lets robot actions share the same interface as text, unlocking transfer from internet-pretrained vision-language models, but coarse binning limits precision, driving research into learned and compressed action codes. ### Activation Function Category: Robot Learning URL: https://almond.bot/glossary/activation-function An activation function is the elementwise nonlinearity applied to a neural network layer's output, enabling the network to represent functions beyond linear maps. ReLU and its variants dominate convolutional and policy networks for their cheap computation and stable gradients, while GELU and SwiGLU are standard in transformer backbones. Saturating functions like tanh are still used to bound continuous action outputs to a normalized range. Why it matters for physical AI: Bounded output activations keep policy commands within actuator limits, and activation choice affects the smoothness of learned control signals, which translates directly into motor wear and motion quality on hardware. ### Active Compliance Category: Control URL: https://almond.bot/glossary/active-compliance Active compliance is robot behavior in which software control, rather than passive mechanical elements, makes a manipulator yield in response to contact forces. It is implemented with force/torque or joint-torque feedback through schemes such as impedance and admittance control, letting a stiff mechanism emulate a spring-damper of programmable stiffness. This contrasts with passive compliance from springs, elastomers, or series elastic actuators, which reacts instantly but cannot be reprogrammed. Why it matters for physical AI: Contact-rich tasks like insertion, wiping, and handovers require yielding to unmodeled contact, and actively compliant control gives learned policies a forgiving action interface that absorbs their prediction errors safely. ### Active Perception Category: Perception URL: https://almond.bot/glossary/active-perception Active perception is the strategy of controlling a robot's sensors or body specifically to improve what it perceives, rather than passively processing whatever arrives. Examples include moving a wrist camera to disambiguate object pose, shifting gaze to reduce occlusion, and next-best-view planning for 3D reconstruction. Formalized by Bajcsy and colleagues in the 1980s, it treats perception as a closed-loop decision problem coupling sensing and action. Why it matters for physical AI: Deployed manipulation systems constantly face occlusion and ambiguity, and policies that learn to reposition cameras or peek before grasping recover information that no amount of passive network capacity can supply. ### Actor-Critic Methods Category: Robot Learning URL: https://almond.bot/glossary/actor-critic-methods Actor-critic methods are a family of reinforcement learning algorithms that maintain two learned components: an actor, the policy that selects actions, and a critic, a value function that evaluates them. The critic's estimates reduce the variance of policy gradient updates relative to pure Monte Carlo returns. Modern instances such as PPO, SAC, and TD3 dominate continuous-control robotics, powering most simulation-trained locomotion and dexterous manipulation results. Why it matters for physical AI: Nearly every large-scale simulation training pipeline for legged robots and dexterous hands rests on an actor-critic algorithm, making their stability and sample efficiency a direct constraint on what behaviors can be trained. ### Actuator Category: Hardware URL: https://almond.bot/glossary/actuator An actuator is a device that converts stored or supplied energy, most commonly electrical, hydraulic, or pneumatic, into mechanical motion. In robots, actuators drive every joint and gripper; the dominant modern choice is an electric motor paired with a gearbox, encoder, and driver electronics, often packaged as an integrated joint module. Key specifications include peak and continuous torque, speed, backdrivability, transparency, and torque density. Why it matters for physical AI: Actuator bandwidth, torque density, and backdrivability set hard physical limits on what any policy can achieve, and proprioceptive actuator designs enable the force sensing that contact-rich learned manipulation depends on. ### Adam Optimizer Category: Robot Learning URL: https://almond.bot/glossary/adam-optimizer Adam is a stochastic gradient descent optimizer, introduced by Kingma and Ba in 2014, that adapts a separate learning rate for each parameter using exponential moving averages of the gradient and its square. Its robustness to hyperparameter choice made it the default optimizer for deep learning, and the decoupled weight decay variant AdamW is standard for training transformers, including robot policy networks and vision-language-action models. Why it matters for physical AI: Virtually every robot policy trained today, from behavior-cloned manipulation networks to large vision-language-action models, is optimized with Adam or AdamW, so its convergence behavior quietly shapes the entire field's results. ### Adaptive Control Category: Control URL: https://almond.bot/glossary/adaptive-control Adaptive control is a class of control techniques in which controller parameters are adjusted online to cope with unknown or time-varying plant dynamics, such as an unmeasured payload or changing friction. Classical schemes include model reference adaptive control (MRAC) and self-tuning regulators, with stability typically argued via Lyapunov methods. It differs from robust control, which fixes a single controller designed to tolerate bounded uncertainty. Why it matters for physical AI: Robots in unstructured settings constantly pick up objects of unknown mass and wear over time; adaptive mechanisms, whether classical or implemented as in-context adaptation in learned policies, keep performance from degrading. ### Adaptive Gripper Category: Hardware URL: https://almond.bot/glossary/adaptive-gripper An adaptive gripper is an end effector whose fingers passively conform to the shape of the grasped object, typically via underactuated linkages, tendons, or compliant materials that distribute contact forces without requiring per-joint control. Designs like the Robotiq 2F series and soft pneumatic grippers can secure a wide range of object geometries with a single close command. This mechanical intelligence trades precise fingertip control for robustness and simplicity. Why it matters for physical AI: Passive conformity absorbs perception and policy errors at the moment of contact, letting learned grasping systems succeed with centimeter-level pose estimates that would defeat a rigid parallel-jaw gripper. ### Admittance Control Category: Control URL: https://almond.bot/glossary/admittance-control Admittance control is an interaction control scheme that measures external forces and responds by modifying the robot's commanded motion, rendering a virtual mass-spring-damper: force in, motion out. It is the dual of impedance control, which commands forces in response to measured motion. Admittance control suits stiff, position-controlled industrial arms with a wrist force/torque sensor, though it can become unstable in contact with very rigid environments. Why it matters for physical AI: Admittance loops let stiff industrial arms be hand-guided for kinesthetic teaching and comply safely during contact-rich learned tasks, making them a common substrate under demonstration collection and deployed manipulation policies. ### Adversarial Robustness Category: Robot Learning URL: https://almond.bot/glossary/adversarial-robustness Adversarial robustness is a model's ability to maintain correct behavior under inputs deliberately crafted or perturbed to cause failure, from imperceptible pixel perturbations to physical patches placed in a scene. Robotics extends the threat model beyond images to state estimates, LiDAR returns, and language instructions given to instruction-following policies. Defenses include adversarial training, input smoothing, randomization, and certified bounds, each trading nominal performance for worst-case guarantees. Why it matters for physical AI: Robots act on their perception with physical consequences, so an adversarial sticker on a stop sign or workpiece is a safety incident rather than a benchmark curiosity, motivating robustness evaluation before deployment. ### Aerial Manipulation Category: Manipulation URL: https://almond.bot/glossary/aerial-manipulation Aerial manipulation is the use of flying robots, typically multirotors equipped with grippers or arms, to grasp, place, or exert forces on objects rather than merely observe them. The core difficulty is dynamic coupling: reaction forces from the arm and contact disturb the floating base, which unlike a fixed manipulator cannot brace against the ground. Applications include infrastructure inspection with contact probes, canopy sampling, and drone-based pick-and-place. Why it matters for physical AI: Floating-base contact tasks stress-test whole-body control and force-aware policies under dynamics far less forgiving than tabletop settings, pushing learned controllers toward tighter perception-action loops. ### Aerospace Robot Category: Applications URL: https://almond.bot/glossary/aerospace-robot An aerospace robot is a robotic system built for aerospace manufacturing or space operations, spanning large-envelope drilling and riveting arms for airframes, automated fiber placement machines for composites, and space systems such as the Canadarm2 manipulator on the ISS and planetary rovers like Perseverance. The domain imposes extreme requirements: tight tolerances over very large parts, radiation-hardened electronics, and operation under multi-second to multi-minute communication delays. Why it matters for physical AI: Long communication latencies make space robotics a proving ground for onboard autonomy, and aerospace assembly's precision-at-scale demands push the limits of calibration, metrology, and learned compensation. ### Affordance Category: Perception URL: https://almond.bot/glossary/affordance An affordance is an action possibility that an object or environment offers an agent, a concept introduced by psychologist James J. Gibson: a handle affords pulling, a flat surface affords placing. In robot learning, affordance prediction typically means estimating dense, pixel- or point-level maps of where and how an object can be grasped, pushed, or articulated, providing an action-centric intermediate representation between raw perception and motor commands. Why it matters for physical AI: Affordance representations transfer across object instances and embodiments better than raw trajectories, giving foundation models a compact interface for grounding language commands like "open the drawer" into actionable regions. ### AgiBot World Category: Data & Benchmarks URL: https://almond.bot/glossary/agibot-world AgiBot World is a large-scale robot manipulation dataset and benchmark released by AgiBot (Zhiyuan Robotics) with academic partners in 2025, containing over a million teleoperated trajectories collected on a homogeneous fleet of more than a hundred dual-arm humanoid robots. It covers hundreds of tasks across domestic, retail, and industrial scenarios with long-horizon and dexterous skills, and was released alongside baseline generalist policies trained on the corpus. Why it matters for physical AI: Fleet-scale, single-embodiment corpora like AgiBot World test whether sheer homogeneous data volume can rival diverse cross-embodiment mixtures, informing the data strategy debates behind current robot foundation models. ### Agile Locomotion Category: Locomotion URL: https://almond.bot/glossary/agile-locomotion Agile locomotion is legged robot movement that approaches the dynamism of animals: running, jumping, vaulting obstacles, rapid direction changes, and recovery from large disturbances. It requires exploiting rather than avoiding the robot's full dynamics, operating near actuator torque and speed limits with flight phases and impacts. Landmark demonstrations include MIT Cheetah's bounding and backflips, Boston Dynamics Atlas parkour, and reinforcement-learned parkour policies on quadrupeds such as ANYmal. Why it matters for physical AI: Agility is where learned controllers have most clearly surpassed hand-designed ones, with sim-trained policies transferring dynamic skills that model-based methods struggled to produce, validating the simulation-to-reality pipeline at the edge of hardware limits. ### Agricultural Robot Category: Applications URL: https://almond.bot/glossary/agricultural-robot An agricultural robot is an autonomous or semi-autonomous machine performing farm tasks such as harvesting, weeding, seeding, spraying, monitoring, and milking. Examples range from GPS-guided autonomous tractors to strawberry-picking manipulators and laser-weeding platforms. The domain combines hard perception problems, including occluded fruit, variable lighting, and deformable plants, with harsh outdoor operation and strong economic pressure from seasonal labor shortages. Why it matters for physical AI: Agriculture offers massive task repetition with high natural variability, an ideal wedge for perception-driven manipulation where each crop's tolerance for slow, imperfect picking exceeds that of industrial assembly lines. ### AI2-THOR Category: Simulation URL: https://almond.bot/glossary/ai2-thor AI2-THOR is an open-source interactive 3D household simulator from the Allen Institute for AI, providing photorealistic indoor scenes where embodied agents can navigate, pick up objects, open containers, and change object states. Built on Unity, it underlies benchmarks including RoomR, RoboTHOR, and the ALFRED language-instruction benchmark, and supports procedural scene generation through ProcTHOR for large-scale training. Why it matters for physical AI: Household simulators such as AI2-THOR supply cheap, resettable environments for training and evaluating instruction-following embodied agents before any real-robot data is collected, seeding many embodied AI training recipes. ### Allegro Hand Category: Hardware URL: https://almond.bot/glossary/allegro-hand The Allegro Hand is a four-fingered, sixteen-degree-of-freedom anthropomorphic robotic hand from Wonik Robotics, with four independently actuated joints per finger. Its relative affordability and open low-level torque control interface made it a de facto standard research platform for dexterous manipulation, appearing in reinforcement learning work on in-hand reorientation and in tactile sensing research where fingertips are retrofitted with sensors. Why it matters for physical AI: Shared platforms like the Allegro Hand make dexterous manipulation results comparable across labs, and its torque-controllable joints suit the contact-rich reinforcement learning experiments driving in-hand skill progress. ### ALOHA Category: Teleoperation URL: https://almond.bot/glossary/aloha ALOHA (A Low-cost Open-source Hardware system for bimanual teleoperation) is a puppeteering rig introduced by Zhao et al. at Stanford in 2023, pairing two small leader arms that a human moves with two follower arms that mirror them, enabling fine bimanual demonstration collection for roughly twenty thousand dollars. Combined with the ACT algorithm, it showed that tasks like slotting a battery or opening a ziploc bag could be learned from about fifty demonstrations; Mobile ALOHA and ALOHA 2 extended the design. Why it matters for physical AI: ALOHA collapsed the cost barrier for high-quality bimanual demonstration data, catalyzing the wave of teleoperation-plus-imitation systems whose datasets now feed vision-language-action model training. ### Anthropomorphic Hand Category: Hardware URL: https://almond.bot/glossary/anthropomorphic-hand An anthropomorphic hand is a robotic end effector that mimics the structure of the human hand, with multiple articulated fingers and often an opposable thumb, ranging from the fully actuated, twenty-plus degree-of-freedom Shadow Hand to underactuated tendon-driven designs. The humanlike form factor targets tool use and environments designed for people, at the cost of mechanical complexity, fragility, and difficult control compared with parallel-jaw grippers. Why it matters for physical AI: Humanoid platforms bet that human-shaped hands unlock human tool ecosystems and make human video a directly relevant pretraining source, but their high-dimensional contact dynamics remain a frontier for learned dexterity. ### Antipodal Grasp Category: Manipulation URL: https://almond.bot/glossary/antipodal-grasp An antipodal grasp is a two-contact grasp in which the contact points can exert forces along the line connecting them, in opposite directions, with each force lying inside its contact's friction cone. Under this condition the grasp achieves force closure with just two fingers, making antipodality the standard geometric test in parallel-jaw grasp planning: samplers search object surfaces or point clouds for approximately antipodal pairs. Why it matters for physical AI: Most deployed grasping still uses parallel-jaw grippers, so antipodal analysis underlies the grasp candidates that data-driven detectors like GraspNet-era systems score and rank in warehouses and labs alike. ### ANYmal Category: Hardware URL: https://almond.bot/glossary/anymal ANYmal is a series of quadruped robots developed at ETH Zurich and commercialized by its spin-off ANYbotics, built around series-elastic ANYdrive joint actuators and targeted at industrial inspection of sites like offshore platforms and chemical plants. It has been the primary hardware platform for influential legged-locomotion research, including Hwangbo et al.'s and Lee et al.'s Science Robotics papers on reinforcement-learned locomotion over rough terrain. Why it matters for physical AI: ANYmal results established that simulation-trained neural controllers could outperform model-based locomotion in the field, providing the template of massively parallel training plus domain randomization now used across legged robotics. ### Apple Vision Pro Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/apple-vision-pro-teleoperation Apple Vision Pro teleoperation is the use of Apple's mixed-reality headset as a robot control interface, streaming its built-in head, hand, and wrist tracking to drive robot arms and hands while the operator sees the robot's stereo camera feed for spatial telepresence. Open-source stacks such as Open-TeleVision and Bunny-VisionPro demonstrated low-latency bimanual and dexterous-hand control, making the headset a popular data-collection front end for humanoid platforms. Why it matters for physical AI: Consumer headsets turn high-fidelity hand tracking into a commodity, lowering the cost of collecting dexterous, egocentric demonstrations at the scale robot foundation model training demands. ### Apprenticeship Learning Category: Robot Learning URL: https://almond.bot/glossary/apprenticeship-learning Apprenticeship learning is learning to perform a task from expert demonstrations by recovering the reward function the expert appears to optimize, then computing a policy against it, rather than mimicking actions directly. Formalized by Abbeel and Ng in 2004 via inverse reinforcement learning with feature matching, it can outperform the demonstrator and generalize beyond demonstrated states, at the price of solving repeated reinforcement learning subproblems. Why it matters for physical AI: Reward-recovery approaches address a core weakness of behavior cloning, distribution drift, and their intellectual descendants inform modern preference-based and adversarial imitation methods used to shape robot behavior. ### Approach Direction Category: Manipulation URL: https://almond.bot/glossary/approach-direction An approach direction is the vector along which a gripper or tool moves toward an object during the final phase of a grasp or insertion, usually expressed in the object or grasp frame. Grasp representations commonly encode it together with the gripper orientation and width, and planners constrain it to avoid collisions with clutter, respect kinematic reachability, and align with task requirements such as top-down picking from bins. Why it matters for physical AI: Grasps that score well geometrically can still fail if the approach sweeps through clutter or exceeds the arm's reachable workspace, so approach reasoning is where grasp detection meets motion planning in real systems. ### AprilTag Category: Perception URL: https://almond.bot/glossary/apriltag AprilTag is a fiducial marker system from the APRIL laboratory at the University of Michigan, consisting of black-and-white square tags whose detection yields both a unique identifier and the full six-degree-of-freedom pose of the tag relative to a calibrated camera. Its detector is robust to rotation, lighting change, and partial occlusion, making it a standard tool for camera calibration, ground-truth tracking, and object or robot localization in labs. Why it matters for physical AI: Fiducials provide cheap millimeter-level ground truth for evaluating learned policies and calibrating camera-to-robot transforms, and their gradual removal from a pipeline is a common milestone toward markerless deployed perception. ### Articulated Object Manipulation Category: Manipulation URL: https://almond.bot/glossary/articulated-object-manipulation Articulated object manipulation is the manipulation of objects with internal joints, such as doors, drawers, laptops, scissors, and appliances, whose parts move relative to each other along revolute or prismatic axes. Success requires estimating the kinematic structure and joint parameters, then generating motions that respect the resulting constraint, since forcing against a hinge axis stalls or damages hardware. Datasets like PartNet-Mobility and benchmarks built on it standardized study of the problem. Why it matters for physical AI: Homes and workplaces are full of doors, drawers, and lids, so generalizing across unseen articulation types is a prerequisite for useful household and service robots, and a persistent failure mode of purely visual policies. ### Articulated Robot Category: Hardware URL: https://almond.bot/glossary/articulated-robot An articulated robot is a robot arm built as a serial chain of rigid links connected by rotary joints, most commonly with six degrees of freedom so the end effector can reach arbitrary positions and orientations within its workspace, or seven for kinematic redundancy. It is the dominant industrial robot morphology, distinguished from SCARA, delta, Cartesian, and parallel architectures by its large dexterous workspace relative to footprint. Why it matters for physical AI: Six- and seven-axis arms are the default embodiment for manipulation research and deployment alike, so their kinematics, singularities, and joint limits implicitly shape the action spaces of most learned manipulation policies. ### Artificial Potential Field Category: Navigation & SLAM URL: https://almond.bot/glossary/artificial-potential-field An artificial potential field is a reactive motion generation method, introduced by Khatib in 1986, that treats the goal as an attractive potential and obstacles as repulsive potentials, moving the robot along the negative gradient of their sum. It is computationally trivial and naturally real-time, but suffers from local minima where attractive and repulsive forces cancel, and from oscillation in narrow passages, so it is typically used as a local layer beneath a global planner. Why it matters for physical AI: Potential-field ideas persist inside modern systems as reactive collision avoidance layers and safety filters that wrap learned policies, providing an interpretable last line of defense near obstacles. ### ArUco Markers Category: Perception URL: https://almond.bot/glossary/aruco-markers ArUco markers are square binary fiducial markers, originating from a University of Cordoba project and integrated into OpenCV, that encode an identifier in an inner bit grid surrounded by a black border. Detecting a marker of known physical size with a calibrated camera yields its six-degree-of-freedom pose, and boards of multiple markers, including ChArUco chessboard hybrids, improve accuracy for camera calibration and workspace registration. Why it matters for physical AI: Because ArUco ships with OpenCV, it is often the fastest route to reliable pose ground truth when setting up demonstration rigs, aligning multi-camera systems, or validating a learned policy's spatial accuracy. ### ASIMO Category: Hardware URL: https://almond.bot/glossary/asimo ASIMO is a humanoid robot developed by Honda and unveiled in 2000, the culmination of a research line begun in the 1980s with the E- and P-series prototypes. Standing about 130 centimeters tall, it demonstrated stable bipedal walking, stair climbing, running, and basic object handling using zero-moment-point-based control, and became for two decades the public face of humanoid robotics before Honda ended development in the late 2010s. Why it matters for physical AI: ASIMO defined the model-based era of humanoid control whose limitations, careful pre-programmed motion and fragility outside structured demos, motivated today's learning-driven approaches to general-purpose humanoids. ### Assistive Robotics Category: Applications URL: https://almond.bot/glossary/assistive-robotics Assistive robotics is the branch of robotics devoted to helping people with disabilities, injuries, or age-related limitations perform daily activities, encompassing wheelchair-mounted manipulators, robot-assisted feeding and dressing, exoskeletons for mobility, and socially assistive robots for therapy and companionship. Research emphasizes safety in close physical contact, personalization to individual users' bodies and preferences, and shared-control interfaces that blend user intent with robot autonomy. Why it matters for physical AI: Care tasks are intimate, contact-rich, and unforgiving of errors, concentrating the hardest problems in safe physical human-robot interaction while demographic aging makes the application economically urgent. ### Asynchronous Inference Category: Software & Middleware URL: https://almond.bot/glossary/asynchronous-inference Asynchronous inference is a robot control architecture in which policy inference runs concurrently with action execution rather than blocking the control loop: while the robot executes the current action chunk, the next chunk is already being computed from more recent observations. Real-time chunking approaches then stitch consecutive chunks to preserve continuity. This decouples a model's inference latency, potentially hundreds of milliseconds for large vision-language-action models, from the robot's control rate. Why it matters for physical AI: Foundation-scale policies are too slow to run synchronously at control rates, so asynchronous execution with chunk stitching is what makes billion-parameter models viable on physical robots without jerky stop-and-go motion. ### Atlas Category: Hardware URL: https://almond.bot/glossary/atlas Atlas is a humanoid robot developed by Boston Dynamics, first revealed in 2013 as a hydraulic platform for the DARPA Robotics Challenge and later famous for parkour, backflips, and dynamic manipulation demonstrations driven by model predictive control. In 2024 Boston Dynamics retired the hydraulic version and introduced a fully electric Atlas with an extended range of joint motion, repositioning the platform toward practical industrial manipulation and learning-based control. Why it matters for physical AI: Atlas demonstrations repeatedly reset public and technical expectations for whole-body dynamism, and the electric redesign signals the industry-wide shift from choreographed model-based motion toward learned, task-general humanoid behavior. ### Attention Mechanism Category: Foundation Models URL: https://almond.bot/glossary/attention-mechanism An attention mechanism is a neural network operation that computes a weighted combination of a set of value vectors, with weights derived from the compatibility between a query and corresponding keys, letting the model dynamically focus on relevant inputs. Scaled dot-product self-attention, introduced with the transformer by Vaswani et al. in 2017, replaced recurrence as the dominant sequence-modeling primitive and underlies essentially all modern language, vision, and robot policy backbones. Why it matters for physical AI: Attention is the mechanism through which vision-language-action models bind language instructions to image regions and past observations to current actions, and its quadratic cost drives the efficiency work needed for onboard deployment. ### Autoencoder Category: Robot Learning URL: https://almond.bot/glossary/autoencoder An autoencoder is a neural network trained to reconstruct its input through a lower-dimensional or otherwise constrained bottleneck, forcing the encoder to learn a compressed latent representation. Variational autoencoders (VAEs) make the latent space probabilistic and generative, and are widely used in robotics for learning latent skill and action spaces, as in the conditional VAE structure of ACT, and as image tokenizers inside world models. Why it matters for physical AI: Latent representations from autoencoders compress high-dimensional observations and demonstrations into spaces where policies and world models train efficiently, a recurring ingredient in modern visuomotor learning stacks. ### Automated Guided Vehicle (AGV) Category: Applications URL: https://almond.bot/glossary/automated-guided-vehicle An automated guided vehicle is a mobile industrial vehicle that follows fixed guidance infrastructure, such as embedded floor wires, magnetic tape, or reflector-based laser triangulation, to transport materials along predetermined routes in factories and warehouses. AGVs stop when obstacles block their path rather than replanning around them, which distinguishes them from autonomous mobile robots (AMRs) that localize and navigate freely using onboard sensing. Why it matters for physical AI: The industry-wide migration from infrastructure-bound AGVs to perception-driven AMRs illustrates the broader arc of physical AI: replacing engineered environmental structure with onboard intelligence that adapts to unmodified spaces. ### Automated Storage and Retrieval System (AS/RS) Category: Applications URL: https://almond.bot/glossary/automated-storage-and-retrieval-system An automated storage and retrieval system is warehouse infrastructure that stores and fetches inventory without manual handling, using stacker cranes running in racking aisles, multi-level shuttle systems, or grid-based robot swarms such as AutoStore's cube storage. AS/RS installations maximize storage density and throughput for goods-to-person fulfillment, delivering totes to stations where humans or robotic arms pick individual items. Why it matters for physical AI: Goods-to-person systems concentrate the remaining manual work into piece-picking at stations, exactly the unstructured grasping problem that learned manipulation is now being deployed to close. ### Autonomous Mobile Robot (AMR) Category: Applications URL: https://almond.bot/glossary/autonomous-mobile-robot An autonomous mobile robot is a wheeled or legged platform that navigates freely using onboard localization, mapping, and path planning rather than fixed guidance infrastructure, replanning around people and obstacles in real time. Typical AMRs fuse LiDAR, cameras, and odometry for SLAM-based navigation in warehouses, hospitals, and factories, with fleet management software coordinating hundreds of units; examples include Locus and MiR platforms and Amazon's warehouse fleets. Why it matters for physical AI: AMRs are among the most widely deployed autonomous systems in commerce, and their mature navigation stacks form the mobility layer onto which learned manipulation is being added to create useful mobile manipulators. ### Autonomous Vehicle (AV) Category: Applications URL: https://almond.bot/glossary/autonomous-vehicle An autonomous vehicle is a car or truck capable of driving with reduced or no human supervision, conventionally graded by the SAE J3016 levels from driver assistance (Level 1-2) through conditional autonomy (Level 3) to full self-driving within an operational design domain (Level 4-5). Production systems combine cameras, radar, and often LiDAR with HD maps, learned perception and prediction, and increasingly end-to-end learned driving policies; Waymo operates commercial Level 4 robotaxi services. Why it matters for physical AI: Self-driving pioneered the playbook now applied to robot manipulation at scale: fleet data engines, simulation-heavy validation, end-to-end learned policies, and safety cases for learned components in the physical world. ### Autonomy Stack Category: Software & Middleware URL: https://almond.bot/glossary/autonomy-stack An autonomy stack is the layered software architecture of an autonomous robot, conventionally decomposed into perception, localization and mapping, prediction, planning, and control modules connected by well-defined interfaces, typically over middleware such as ROS 2. The modular design permits independent development, testing, and validation of each layer, in contrast to end-to-end learned systems that map sensors to actions with a single network, and hybrid stacks that embed learned modules within the classical decomposition. Why it matters for physical AI: Whether foundation policies replace the full stack or slot in as modules is a central architectural question for deployment, since modular interfaces provide the testability and traceability that safety cases currently require. ### Autoregressive Policy Category: Foundation Models URL: https://almond.bot/glossary/autoregressive-policy An autoregressive policy is a robot policy that generates its output as a sequence of discrete tokens, each conditioned on previously generated tokens, exactly as a language model generates text. Actions are discretized via binning or a learned tokenizer, and the policy factorizes the action distribution across dimensions and timesteps; RT-1, RT-2, and OpenVLA are prominent examples. Sequential decoding limits inference speed, motivating parallel-decoding and diffusion-based alternatives. Why it matters for physical AI: Casting control as next-token prediction lets robot policies inherit the architectures, pretraining, and scaling behavior of large language models, though decoding latency remains the key obstacle to high-rate control. ### AutoRT Category: Foundation Models URL: https://almond.bot/glossary/autort AutoRT is a system from Google DeepMind, presented in 2024, that uses foundation models to orchestrate autonomous data collection across a fleet of mobile manipulators. A vision-language model describes each robot's scene, a large language model proposes candidate tasks and filters them through a written constitution of safety and feasibility rules, and tasks are then executed by teleoperators, scripted policies, or learned policies such as RT-2. It coordinated over twenty robots collecting tens of thousands of episodes. Why it matters for physical AI: Scaling robot learning requires data engines, not just models, and AutoRT prototyped how language models can supervise fleets to gather diverse real-world experience with minimal human oversight. ## Glossary: B ### B-spline Category: Math & Kinematics URL: https://almond.bot/glossary/b-spline A B-spline is a piecewise polynomial curve defined as a weighted combination of control points using basis functions of a chosen degree over a knot vector, providing local control: moving one control point reshapes only a neighborhood of the curve. B-splines of degree k are k-1 times continuously differentiable at simple knots, which makes them a standard representation for smooth robot joint trajectories and Cartesian paths in trajectory optimization. Why it matters for physical AI: Smooth, locally editable trajectory parameterizations keep accelerations and jerks within actuator limits, and optimizing over a handful of control points is far cheaper than optimizing every waypoint of a dense trajectory. ### Backdrivable Actuator Category: Hardware URL: https://almond.bot/glossary/backdrivable-actuator A backdrivable actuator is an actuator whose output can be moved by external forces without damage and with low resistance, allowing the mechanism to be pushed by hand or by contact with the environment. Backdrivability generally requires low gear ratios, as in quasi-direct-drive designs popularized by the MIT Cheetah, since high-ratio gearboxes like harmonic drives reflect large apparent inertia and friction to the output. Why it matters for physical AI: Backdrivable joints let robots sense contact through motor currents and yield safely on impact, enabling proprioceptive force control and kinesthetic teaching without dedicated force sensors, and reducing injury risk around people. ### Background Subtraction Category: Perception URL: https://almond.bot/glossary/background-subtraction Background subtraction is a computer vision technique that segments moving or novel objects in a scene by maintaining a statistical model of the static background and flagging pixels that deviate from it. Classical methods include Gaussian mixture models per pixel and kernel density approaches, which handle gradual illumination change but struggle with camera motion and dynamic backgrounds. Deep segmentation models have largely replaced it where compute allows. Why it matters for physical AI: Lightweight change detection still earns its keep on embedded robot platforms for workspace monitoring, safety curtains, and triggering more expensive perception only when something in the scene actually changes. ### Backlash Category: Hardware URL: https://almond.bot/glossary/backlash Backlash is the mechanical play between mating gear teeth or transmission elements, manifesting as a small dead zone where the output does not move when the input reverses direction. It introduces hysteresis and nonlinearity that degrade positioning accuracy and can cause limit cycles in feedback control. Zero-backlash transmissions such as harmonic drives, preloaded ball screws, and split-gear arrangements mitigate it at added cost and friction. Why it matters for physical AI: Unmodeled backlash corrupts the assumption that commanded and actual motion match, injecting noise into demonstration data and blurring the fine positioning that learned insertion and assembly policies must achieve. ### Backpropagation Category: Robot Learning URL: https://almond.bot/glossary/backpropagation Backpropagation is the algorithm for computing gradients of a loss function with respect to every parameter of a neural network by applying the chain rule backward through the computation graph, layer by layer. Popularized for training multilayer networks by Rumelhart, Hinton, and Williams in 1986, it is the engine of essentially all deep learning, implemented as reverse-mode automatic differentiation in frameworks like PyTorch and JAX. Why it matters for physical AI: Every learned robot policy, from grasp detectors to vision-language-action models, is trained by backpropagation, and differentiable simulators extend the same machinery to computing gradients through physics itself. ### Backstepping Category: Control URL: https://almond.bot/glossary/backstepping Backstepping is a recursive nonlinear control design technique for systems in strict-feedback form, in which a stabilizing virtual control law is designed for an inner subsystem and then extended outward one integrator at a time, with a Lyapunov function constructed at each step to guarantee stability. It is widely applied to quadrotor flight control, underactuated marine vehicles, and manipulator tracking, and combines naturally with adaptive parameter estimation. Why it matters for physical AI: Provably stable nonlinear controllers remain the trusted inner loops around which learned outer policies are wrapped, and backstepping supplies such guarantees for the cascaded structure typical of drones and vehicles. ### Balance Recovery Category: Locomotion URL: https://almond.bot/glossary/balance-recovery Balance recovery is the set of strategies a legged robot uses to avoid falling after a push, slip, or trip, conventionally categorized as ankle, hip, and stepping strategies in order of increasing disturbance magnitude, with capture point theory formalizing where a foot must be placed to come to rest. Reinforcement-learned controllers now discover recovery behaviors, including multi-step corrections and even fall-and-stand-up sequences, that exceed hand-designed strategies. Why it matters for physical AI: Push recovery determines whether a humanoid survives contact with the messy world of bumped tables and shifting loads, making it a gating capability for deploying legged robots outside labs. ### Ball Bearing Category: Hardware URL: https://almond.bot/glossary/ball-bearing A ball bearing is a rolling-element bearing that supports rotating shafts by interposing hardened steel or ceramic balls between inner and outer races, replacing sliding friction with much lower rolling friction. Variants include deep-groove bearings for radial loads, angular-contact pairs for combined loads, and thin-section and crossed-roller types used in robot joints where space is tight and moment stiffness matters. Why it matters for physical AI: Joint stiffness, friction, and lifetime trace back to bearing selection, and bearing friction is part of the unmodeled dynamics that simulation-trained policies must be robust to when transferred onto real hardware. ### Ball Screw Category: Hardware URL: https://almond.bot/glossary/ball-screw A ball screw is a linear actuation mechanism that converts rotary motion to linear motion via a threaded shaft and a nut containing recirculating balls, achieving efficiencies above ninety percent compared with roughly a quarter to half for sliding lead screws. Preloaded nuts eliminate backlash, giving the precision and stiffness required in CNC machine tools, humanoid leg linear actuators, and high-force linear axes. Why it matters for physical AI: Several current humanoid designs use ball-screw linear actuators for high force density and impact robustness in legs, a hardware choice that shapes the torque bandwidth available to learned locomotion controllers. ### Base Frame Category: Math & Kinematics URL: https://almond.bot/glossary/base-frame A base frame is the coordinate frame fixed to a robot's base link, serving as the root reference in which forward kinematics expresses end-effector poses and in which workspace targets are commonly specified. Frame conventions and transforms between base, world, camera, and tool frames are managed explicitly in software, for example via the tf2 library in ROS, and errors in these transforms are a classic source of systematic positioning bias. Why it matters for physical AI: Learned policies trained with observations or actions in one frame silently fail when deployed with mismatched frame conventions, making rigorous frame bookkeeping a mundane but critical part of policy portability. ### Base Policy Category: Foundation Models URL: https://almond.bot/glossary/base-policy A base policy is a generalist pretrained robot policy that serves as the starting point for adaptation, analogous to a base model in language modeling: it is trained on broad multi-task, often multi-embodiment data, then specialized to a target robot or task via fine-tuning, distillation, or steering. The term also denotes the prior policy that residual reinforcement learning or test-time refinement methods improve upon. Why it matters for physical AI: The pretrain-then-adapt recipe reduces per-task data needs from thousands of demonstrations to tens, making base policy quality and coverage the central asset of robot foundation model efforts. ### Batch Normalization Category: Robot Learning URL: https://almond.bot/glossary/batch-normalization Batch normalization is a neural network layer, introduced by Ioffe and Szegedy in 2015, that normalizes activations using the mean and variance of the current mini-batch, then rescales them with learned parameters, stabilizing and accelerating training of deep networks. Its dependence on batch statistics complicates recurrent models, small batches, and reinforcement learning, so transformers and many policy networks favor layer normalization or group normalization instead. Why it matters for physical AI: Normalization layers whose behavior differs between training and inference are a recurring source of sim-to-real and deployment bugs in visuomotor policies, motivating the field's drift toward batch-independent alternatives. ### Batch Size Category: Robot Learning URL: https://almond.bot/glossary/batch-size Batch size is the number of training examples processed together in one forward-backward pass before a gradient update, balancing gradient estimate variance against memory use and hardware utilization. Large-scale policy training aggregates batches across many accelerators, while on-policy reinforcement learning in simulation gathers batches from thousands of parallel environments. Learning rate and batch size interact strongly, with linear or square-root scaling rules commonly applied. Why it matters for physical AI: Massively parallel simulation made huge on-policy batches cheap, a key enabler of learned locomotion, while vision-language-action pretraining inherits the large-batch optimization recipes of language model training. ### Battery Swapping Category: Hardware URL: https://almond.bot/glossary/battery-swapping Battery swapping is the practice of exchanging a robot's depleted battery pack for a charged one, manually or at automated swap stations, so the platform returns to service in minutes rather than waiting through a charge cycle. It is common for drone fleets, warehouse AMRs, and inspection robots, and several humanoid manufacturers have adopted hot-swappable packs; the alternative approach is opportunistic autonomous dock charging. Why it matters for physical AI: Fleet economics hinge on utilization, and swap-capable designs keep robots earning around the clock, while energy budgets remain a binding constraint on the compute that can run onboard for large learned models. ### Baxter Category: Hardware URL: https://almond.bot/glossary/baxter Baxter is a dual-arm collaborative robot introduced by Rethink Robotics, the company co-founded by Rodney Brooks, in 2012, featuring two seven-degree-of-freedom arms with series elastic actuators, a screen face for conveying state, and demonstration-by-guiding programming aimed at small manufacturers. Commercially it struggled against faster, more precise competitors and Rethink closed in 2018, but Baxter and its one-armed sibling Sawyer became fixtures of robot learning research labs. Why it matters for physical AI: Baxter mainstreamed compliant, human-safe arms and hand-guided teaching, and as an affordable bimanual research platform it hosted a generation of early deep robot learning experiments before today's low-cost arms arrived. ### Bayesian Optimization Category: Robot Learning URL: https://almond.bot/glossary/bayesian-optimization Bayesian optimization is a sample-efficient method for optimizing expensive black-box functions by fitting a probabilistic surrogate, typically a Gaussian process, to observed evaluations and choosing the next query point via an acquisition function that balances exploration and exploitation. In robotics it is used to tune controller gains, hardware morphology, and simulation parameters directly from physical trials, where each evaluation costs real robot time. Why it matters for physical AI: When each experiment consumes minutes of robot wear and human supervision, extracting the most from a handful of trials matters, and Bayesian optimization remains the standard tool for hardware-in-the-loop tuning. ### BC-Z Category: Robot Learning URL: https://almond.bot/glossary/bc-z BC-Z is a large-scale imitation learning system from Google, presented by Jang et al. in 2022, that trained a single policy on over 25,000 teleoperated and shared-autonomy demonstrations spanning 100 manipulation tasks, conditioning on either a language command or a video of a human performing the task. It demonstrated zero-shot generalization to held-out tasks by composing the conditioning space, prefiguring the language-conditioned generalist policies that followed. Why it matters for physical AI: BC-Z provided early evidence that scaling demonstration diversity yields task-level generalization, helping justify the data-scaling strategy behind subsequent robot foundation models such as RT-1 and RT-2. ### BEHAVIOR Benchmark Category: Data & Benchmarks URL: https://almond.bot/glossary/behavior-benchmark BEHAVIOR is a benchmark family from Stanford for embodied AI in household settings, whose current form, BEHAVIOR-1K, defines 1,000 everyday activities such as cleaning, cooking, and tidying, grounded in surveys of what people most want automated. Tasks are specified in a symbolic predicate language and simulated in OmniGibson atop NVIDIA Omniverse, with realistic homes, articulated objects, and fluid and thermal effects, and long-horizon mobile manipulation remains largely unsolved on it. Why it matters for physical AI: Household benchmarks with rigorous success predicates expose how far current policies are from useful home autonomy, providing a common yardstick for long-horizon mobile manipulation progress. ### Behavior Cloning (BC) Category: Robot Learning URL: https://almond.bot/glossary/behavior-cloning Behavior cloning is the simplest form of imitation learning: a policy is trained by supervised learning to map observations to the actions an expert took, treating demonstrations as a labeled dataset. Its central weakness is compounding error, since small mistakes drift the robot into states absent from training data, formalized by Ross et al.'s DAgger analysis. Modern remedies include action chunking, expressive generative action heads, and simply scaling data, which together power systems from ACT to vision-language-action models. Why it matters for physical AI: Nearly all current robot foundation models are trained substantially by behavior cloning on teleoperated demonstrations, so its data appetite and distribution-shift failure modes define the field's core scaling economics. ### Behavior Foundation Model (BFM) Category: Foundation Models URL: https://almond.bot/glossary/behavior-foundation-model A behavior foundation model is a pretrained model of motor behavior intended to serve as a general substrate for many downstream tasks, typically trained on large corpora of motion data such as human motion capture or diverse robot trajectories. In humanoid research, BFMs learn latent skill spaces for whole-body control that can be steered by tracking targets, rewards, or language, so new tasks require only lightweight steering rather than training a controller from scratch. Why it matters for physical AI: Reusable motor priors could do for whole-body control what language model pretraining did for text, cutting per-task training from weeks of reinforcement learning to prompt-like specification of goals. ### Behavior Transformer (BeT) Category: Robot Learning URL: https://almond.bot/glossary/behavior-transformer Behavior Transformer is an imitation learning method, introduced by Shafiullah et al. in 2022, designed to clone unlabeled, multimodal demonstration data containing several distinct ways of performing tasks. It discretizes actions with k-means clustering, has a transformer predict the cluster plus a continuous residual offset, and thereby captures multiple behavior modes that mean-regression policies average away. VQ-BeT later replaced the clustering with vector-quantized codes. Why it matters for physical AI: Human demonstrations are inherently multimodal, and BeT's discretize-plus-offset scheme was an influential early answer to mode collapse, a design thread continued by diffusion and flow-based action heads. ### Behavior Tree Category: Software & Middleware URL: https://almond.bot/glossary/behavior-tree A behavior tree is a hierarchical task-control architecture in which a tree of composite nodes, such as sequences, fallbacks, and parallels, ticks leaf nodes that perform actions or check conditions, returning success, failure, or running status upward. Originating in game AI, behavior trees became standard in robotics through implementations like BehaviorTree.CPP, which orchestrates the ROS 2 Nav2 stack, offering modularity and reactive execution that finite state machines handle poorly. Why it matters for physical AI: Deployed systems still need interpretable orchestration above learned skills, and behavior trees are a dominant pattern for sequencing policies, handling failures, and encoding recovery logic that operators can audit. ### Behavioral Regularization Category: Robot Learning URL: https://almond.bot/glossary/behavioral-regularization Behavioral regularization is a family of techniques, most prominent in offline reinforcement learning, that constrain a learned policy to stay close to the behavior policy that generated the training data, penalizing divergence measured by KL divergence, action-space distance, or implicit weighting. Algorithms such as TD3+BC, AWAC, and advantage-weighted regression instantiate it, preventing the policy from exploiting value-function errors on out-of-distribution actions. Why it matters for physical AI: Improving beyond demonstrations without real-world exploration requires offline reinforcement learning, and behavioral regularization is the mechanism that keeps such improvement from hallucinating actions the data cannot support. ### Belief State Category: Robot Learning URL: https://almond.bot/glossary/belief-state A belief state is a probability distribution over the possible true states of a partially observable system, summarizing everything the observation history implies about the world. It is the sufficient statistic on which optimal policies for partially observable Markov decision processes (POMDPs) operate, and it is maintained in practice by Bayes filters such as Kalman and particle filters, or implicitly by the hidden state of recurrent and transformer policies. Why it matters for physical AI: Robots never observe the world fully, with occlusion, sensor noise, and unknown object properties everywhere, so acting on calibrated uncertainty rather than point estimates separates robust deployed behavior from brittle demos. ### Bellman Equation Category: Robot Learning URL: https://almond.bot/glossary/bellman-equation The Bellman equation is the recursive relationship stating that the value of a state equals the expected immediate reward plus the discounted value of the successor state, formulated by Richard Bellman in the 1950s as the foundation of dynamic programming. Its fixed point defines the optimal value function, and temporal-difference methods including Q-learning and the critics of actor-critic algorithms are constructed as stochastic approximations of Bellman backups. Why it matters for physical AI: Every value-based component in modern robot reinforcement learning, from locomotion critics to offline Q-functions, is trained against Bellman targets, whose approximation errors drive the stability tricks these systems require. ### Belt Drive Category: Hardware URL: https://almond.bot/glossary/belt-drive A belt drive is a power transmission that couples two pulleys with a flexible belt, in robotics almost always a toothed synchronous (timing) belt that prevents slip while allowing the motor to be mounted away from the driven joint. Belts reduce reflected inertia at distal joints, run quietly, tolerate shock loads, and add elasticity, but they stretch over time and require tensioning, trading some stiffness and precision against gears. Why it matters for physical AI: Remote motor placement via belts keeps arm inertia low, improving both dynamic performance and collision safety, a design pattern common in low-cost arms used for learning research and in high-speed pick-and-place hardware. ### BERT Category: Foundation Models URL: https://almond.bot/glossary/bert BERT (Bidirectional Encoder Representations from Transformers) is a language model from Google, introduced by Devlin et al. in 2018, that pretrains a transformer encoder with masked-token and next-sentence objectives, then fine-tunes it for downstream tasks. It established the pretrain-then-fine-tune paradigm at scale, and its bidirectional masked modeling inspired representation learners across domains, including masked autoencoders in vision and BERT-style objectives over robot trajectory tokens. Why it matters for physical AI: The transfer learning recipe BERT proved out, generic pretraining followed by cheap task adaptation, is precisely the template robot foundation models now apply to sensorimotor data. ### Bezier Curve Category: Math & Kinematics URL: https://almond.bot/glossary/bezier-curve A Bezier curve is a parametric polynomial curve defined by a set of control points and blended through Bernstein basis polynomials, beginning at the first control point, ending at the last, and remaining within the control points' convex hull. Popularized by Pierre Bezier for automotive body design, in robotics it parameterizes smooth trajectory segments, swing-foot trajectories in legged locomotion, and lane-change paths in autonomous driving. Why it matters for physical AI: Low-order Bezier segments give planners a compact way to generate smooth, bounded motions whose derivatives are easy to constrain, keeping commanded trajectories within the dynamic limits real actuators can track. ### Bilateral Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/bilateral-teleoperation Bilateral teleoperation is remote robot operation in which information flows both ways: the operator's motions command the remote robot while forces measured or estimated at the remote site are reflected back to the operator's input device as haptic feedback. The scheme creates a closed force loop whose stability degrades with communication delay, studied through passivity theory, wave variables, and time-domain passivity controllers. Why it matters for physical AI: Force feedback lets operators feel contact during demonstration collection, producing richer data for contact-sensitive tasks, and haptic channels are increasingly logged as training signal alongside vision and proprioception. ### Bill of Materials (BOM) Category: Hardware URL: https://almond.bot/glossary/bill-of-materials A bill of materials is the complete structured list of parts, components, and raw materials required to build a product, with quantities, part numbers, suppliers, and costs, often organized hierarchically by subassembly. In robotics, BOM cost is the standard shorthand for hardware affordability, and open-source projects like ALOHA and various low-cost arms publish full BOMs so others can replicate the hardware. Why it matters for physical AI: Falling BOM costs for capable arms and teleoperation rigs democratized demonstration collection, directly expanding the community and data volume behind current robot learning progress. ### Bimanual Manipulation Category: Manipulation URL: https://almond.bot/glossary/bimanual-manipulation Bimanual manipulation is the coordinated use of two robot arms to perform tasks impossible or awkward for one, including stabilizing an object with one hand while the other acts on it, stretching deformables, re-grasping, and handling large items. Coordination modes range from symmetric and leader-follower motions to fully asymmetric role division, and the doubled action space with tight inter-arm constraints makes both planning and learning markedly harder than single-arm settings. Why it matters for physical AI: Most human manual work is two-handed, so bimanual competence gates a large share of real tasks, and systems like ALOHA showed learned policies can master coordination that defeated analytic approaches. ### Bimanual Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/bimanual-teleoperation Bimanual teleoperation is the simultaneous remote control of two robot arms by a single operator, using paired leader arms as in ALOHA, dual haptic devices, motion-capture suits, or hand tracking from VR headsets. It demands interfaces that preserve inter-arm coordination while managing the operator's cognitive load, and it is currently the dominant method for collecting the two-handed demonstration data on which bimanual and humanoid policies train. Why it matters for physical AI: The throughput and ergonomics of two-handed teleoperation set the cost of bimanual training data, making rig design a first-order factor in scaling humanoid manipulation datasets. ### Bin Picking Category: Applications URL: https://almond.bot/glossary/bin-picking Bin picking is the task of extracting individual items from a container of randomly piled, often overlapping objects, a longstanding industrial challenge combining severe occlusion, unknown object poses, gripper-bin collisions, and entangled parts. Modern systems pair 3D sensing with learned grasp detection and are widely deployed in logistics piece-picking; Dex-Net and the Amazon Picking Challenge were influential in moving the field from CAD-matched parts to novel-object picking. Why it matters for physical AI: Warehouse piece-picking is among the first high-volume commercial wins for learned grasping, and its residual failure cases, transparent packaging, entanglement, and deformables, still steer perception and gripper research. ### Bioinspired Robotics Category: Applications URL: https://almond.bot/glossary/bioinspired-robotics Bioinspired robotics is the design of robots that borrow principles from biological organisms, spanning gecko-adhesive climbers, undulating snake robots, flapping-wing micro air vehicles, octopus-like soft arms, and legged machines whose gaits and compliant tendons echo animal biomechanics. Rather than copying anatomy outright, the field abstracts mechanisms such as passive dynamics, distributed compliance, and sensory-motor reflexes that evolution converged on for robustness and efficiency. Why it matters for physical AI: Animals remain the existence proof for agile, efficient, damage-tolerant physical intelligence, and bioinspired principles like compliance and reflex layering repeatedly resurface inside successful learned-control architectures. ### Bipedal Robot Category: Locomotion URL: https://almond.bot/glossary/bipedal-robot A bipedal robot is a legged robot that walks on two legs, trading the static stability of quadrupeds for a human-compatible form factor and footprint. Control has progressed from zero-moment-point walking in machines like ASIMO, through hybrid-dynamics approaches on platforms such as Cassie, to reinforcement-learned controllers now standard on humanoids from Digit to current commercial platforms. The dynamics are underactuated during single support, making balance a continuous active process. Why it matters for physical AI: The humanoid industry's central bet is that two-legged, human-shaped robots can inhabit unmodified human spaces, so bipedal robustness on stairs, clutter, and slips directly gates commercial viability. ### Bitter Lesson Category: Foundation Models URL: https://almond.bot/glossary/bitter-lesson The Bitter Lesson is the argument, articulated by Richard Sutton in a 2019 essay, that seventy years of AI research show general methods leveraging computation, chiefly search and learning, ultimately outperform approaches built on human-designed domain knowledge, which deliver short-term gains but plateau. The essay is invoked constantly in robotics debates over end-to-end learning versus modular pipelines built from human priors about geometry, dynamics, and task structure. Why it matters for physical AI: Whether robotics obeys the Bitter Lesson is the field's live strategic question: data-and-compute scaling drives current foundation model bets, while physical constraints and data scarcity keep structured priors competitive. ### Blind Locomotion Category: Locomotion URL: https://almond.bot/glossary/blind-locomotion Blind locomotion is legged locomotion controlled without exteroceptive sensing such as cameras or LiDAR, relying solely on proprioception: joint positions, torques, and inertial measurements. Lee et al.'s 2020 Science Robotics work showed a proprioceptive policy trained in simulation with a teacher-student scheme could carry ANYmal over mud, snow, and vegetation, feeling out terrain through contact. Perceptive controllers now build on blind policies as a robust fallback layer. Why it matters for physical AI: Cameras fail in dust, darkness, and glare, so a proprioceptive core that survives total perception loss remains the reliability backbone of deployed legged systems. ### Body Schema Category: Perception URL: https://almond.bot/glossary/body-schema A body schema is an agent's internal model of its own body, including link geometry, joint configuration, dynamics, and the sensory consequences of its motion, a concept imported from neuroscience where it explains phenomena like tool incorporation. In robotics it covers learned self-models: networks that predict the robot's occupancy for self-collision checking, calibrate kinematics from experience, and adapt online when the body changes through damage or grasped tools. Why it matters for physical AI: Robots that learn and update their own body models can absorb calibration drift, hardware variation, and tool use without re-engineering, a capability cross-embodiment foundation policies implicitly require. ### Boston Dynamics Spot Category: Hardware URL: https://almond.bot/glossary/boston-dynamics-spot Spot is a quadruped robot from Boston Dynamics, commercially released in 2020 as the company's first mass-produced product, weighing roughly 32 kilograms with electric actuation, stereo perception, an optional arm, and a payload and API ecosystem for third-party integration. Deployed in the thousands for industrial inspection, construction monitoring, and public safety, it is the most widely fielded legged robot and a common research platform via its SDK. Why it matters for physical AI: Spot proved legged robots can survive commercial field duty at fleet scale, and its deployments generate the operational lessons, autonomy gaps, and inspection data shaping industrial legged-robot practice. ### Bounding Box Category: Perception URL: https://almond.bot/glossary/bounding-box A bounding box is a rectangular region, axis-aligned or oriented, in 2D images or as a cuboid in 3D space, that localizes a detected object, forming the standard output representation of object detectors from Faster R-CNN and YOLO to DETR-style transformers. Open-vocabulary detectors now produce boxes for arbitrary text queries, and 3D boxes with position, size, and heading are the lingua franca of autonomous driving perception. Why it matters for physical AI: Boxes are the interface through which detection feeds grasping, tracking, and language grounding in robot pipelines, though their coarseness relative to masks and poses limits fine manipulation, motivating richer representations. ### Bounding Gait Category: Locomotion URL: https://almond.bot/glossary/bounding-gait A bounding gait is a dynamic quadruped gait in which the front leg pair and hind leg pair alternate contact, with the body pitching cyclically and flight phases between touchdowns, as seen in running dogs and rabbits. It stresses actuators with high-impulse impacts and demands precise pitch control, making it a benchmark behavior in dynamic legged robotics from the MIT Cheetah's bounding to reinforcement-learned gait repertoires. Why it matters for physical AI: Mastering high-flight-phase gaits certifies that a control stack can exploit full body dynamics near hardware limits, capabilities that carry over to jumping, gap crossing, and rough-terrain agility in the field. ### BridgeData V2 Category: Data & Benchmarks URL: https://almond.bot/glossary/bridgedata-v2 BridgeData V2 is an open dataset of roughly 60,000 robot manipulation trajectories, over 50,000 of them teleoperated demonstrations plus scripted rollouts, collected on low-cost WidowX 250 arms across 24 environments such as toy kitchens, with natural-language task annotations. Released by Berkeley-led collaborators in 2023 to support scalable multi-task learning research, it became a core training corpus within Open X-Embodiment and for generalist policies including OpenVLA and Octo. Why it matters for physical AI: Bridge established that shared, low-cost-platform datasets let labs worldwide train and compare generalist policies on common ground, seeding the open data ecosystem behind current vision-language-action models. ### Brownfield Deployment Category: Applications URL: https://almond.bot/glossary/brownfield-deployment A brownfield deployment is the introduction of robots or automation into an existing facility built for human workers, as opposed to a greenfield site designed around the automation from the start. Brownfield settings impose the constraints as found: narrow aisles, stairs, legacy equipment, variable lighting, and workflows that cannot pause, so systems must adapt to the environment rather than the reverse. Why it matters for physical AI: The commercial thesis for humanoids and general-purpose robots rests on brownfield economics, where machines flexible enough for unmodified human spaces avoid the capital cost that confines fixed automation to greenfield sites. ### Brushed DC Motor Category: Hardware URL: https://almond.bot/glossary/brushed-dc-motor A brushed DC motor is an electric motor in which mechanical carbon brushes press against a rotating commutator to switch current through the rotor windings, producing torque roughly proportional to current from a simple DC supply. Brushes make the motor cheap and trivially controllable but wear over time, generate electrical noise and friction, and cap speed and power density, which is why brushless designs dominate modern robot joints. Why it matters for physical AI: Brushed motors persist in low-cost educational arms, grippers, and prototypes where simplicity beats longevity, but their friction and wear characteristics inject exactly the unmodeled dynamics that complicate precise learned control. ### Brushless DC Motor (BLDC) Category: Hardware URL: https://almond.bot/glossary/brushless-dc-motor A brushless DC motor is an electric motor that replaces mechanical commutation with electronic switching of stator winding currents, coordinated with rotor position from encoders or Hall sensors, typically via field-oriented control. Eliminating brushes yields higher efficiency, power density, lifetime, and speed, and permits precise torque control through current regulation. Outrunner BLDC designs with high pole counts power the quasi-direct-drive joints of modern legged robots and low-cost arms. Why it matters for physical AI: High-torque-density BLDC modules are the enabling hardware of the current robotics wave, providing the bandwidth and current-based torque sensing that dynamic learned controllers exploit for contact-rich behavior. ### Bullet Physics (PyBullet) Category: Simulation URL: https://almond.bot/glossary/bullet-physics Bullet is an open-source physics engine providing rigid-body dynamics, articulated multibody simulation via Featherstone's algorithm, and collision detection, with PyBullet exposing it to Python along with URDF loading, synthetic cameras, and reinforcement learning environment utilities. Maintained for years by Erwin Coumans, it became a default for robot learning research in the late 2010s before GPU-parallel simulators like Isaac Gym absorbed much large-scale training. Why it matters for physical AI: Free, scriptable physics simulation lowered the barrier to policy training and benchmarking for an entire research generation, and PyBullet remains a common choice for lightweight prototyping and reproducible baselines. ### Bundle Adjustment Category: Navigation & SLAM URL: https://almond.bot/glossary/bundle-adjustment Bundle adjustment is the joint nonlinear least-squares optimization of camera poses and 3D landmark positions to minimize reprojection error across many images, the gold-standard refinement step in structure from motion and visual SLAM. Its sparse structure is exploited via the Schur complement and solvers such as Ceres, g2o, and GTSAM, and keyframe-based systems like ORB-SLAM run local and global bundle adjustment to maintain map consistency. Why it matters for physical AI: Accurate maps and camera poses from bundle adjustment underpin robot localization, 3D reconstruction for digital twins, and the camera-pose ground truth used to train and evaluate spatially aware perception models. ### Byte-Pair Encoding (BPE) Category: Foundation Models URL: https://almond.bot/glossary/byte-pair-encoding Byte-pair encoding is a subword tokenization algorithm that builds a vocabulary by iteratively merging the most frequent pair of adjacent symbols in a corpus, adapted for neural machine translation by Sennrich et al. in 2016 and used by the GPT family. Beyond text, robotics has adopted BPE-style compression for action sequences: the FAST tokenizer applies it after a discrete cosine transform to encode high-frequency action chunks in few tokens. Why it matters for physical AI: Compressed action vocabularies shorten the token sequences vision-language-action models must generate, directly cutting inference latency and making autoregressive policies practical for high-rate dexterous control. ## Glossary: C ### Cable Harness Category: Hardware URL: https://almond.bot/glossary/cable-harness A cable harness is a bundled assembly of wires, connectors, and protective sheathing that routes power and signals through a robot's structure. In articulated robots, harnesses must survive millions of flex cycles across moving joints, so designs use high-flex-rated conductors, strain relief, and service loops. Poor harness routing is a leading cause of intermittent faults and premature failure in deployed robot arms. Why it matters for physical AI: Reliable long-horizon data collection and 24/7 deployment depend on harnesses that tolerate continuous joint motion; cable failures are among the most common hardware faults that interrupt fleet uptime and corrupt teleoperation sessions. ### Cable Manipulation Category: Manipulation URL: https://almond.bot/glossary/cable-manipulation Cable manipulation is the robotic handling of deformable linear objects such as cables, ropes, and wires, including tasks like routing, untangling, knot tying, and connector insertion. Because cables have infinite-dimensional state and complex friction dynamics, classical rigid-body planners fail, and modern approaches rely on learned dynamics models, visual tracking of keypoints, or end-to-end imitation learning. It is a canonical benchmark for deformable object manipulation research. Why it matters for physical AI: Cables appear everywhere in manufacturing, e-commerce, and household settings, and their unpredictable dynamics make them a stress test for learned policies that must generalize beyond rigid objects. ### Calibration Category: Perception URL: https://almond.bot/glossary/calibration Calibration is the process of estimating the parameters that relate a robot's sensors, actuators, and kinematic model to physical reality, such as camera intrinsics, sensor-to-robot transforms, joint offsets, and force sensor biases. Accurate calibration ensures that measurements from different sensors can be fused in a common reference frame and that commanded motions match real-world outcomes. Calibration typically drifts over time due to wear, thermal effects, and mechanical shocks, requiring periodic re-estimation. Why it matters for physical AI: Learned policies trained on one robot's data implicitly bake in its calibration; miscalibrated cameras or kinematics shift the observation distribution and silently degrade policy performance across a fleet. ### CALVIN Category: Data & Benchmarks URL: https://almond.bot/glossary/calvin CALVIN (Composing Actions from Language and Vision) is a simulated benchmark for long-horizon, language-conditioned manipulation, introduced by Mees et al. in 2022. A Franka arm in four tabletop environments must complete chains of up to five sequential instructions such as opening drawers, pushing blocks, and toggling lights, using onboard RGB observations. Its evaluation protocol measures how many consecutive subtasks a policy completes, testing both instruction grounding and long-horizon consistency. Why it matters for physical AI: CALVIN remains a standard yardstick for language-conditioned policies and hierarchical vision-language-action systems, exposing failure modes in instruction chaining that single-task benchmarks miss. ### Camera Calibration Category: Perception URL: https://almond.bot/glossary/camera-calibration Camera calibration is the estimation of a camera's intrinsic parameters (focal length, principal point, lens distortion) and often its extrinsic pose relative to another frame, typically by observing a known pattern such as a checkerboard or ChArUco board from multiple views. Zhang's method (2000) is the standard planar-target algorithm implemented in OpenCV. Calibration enables metric 3D reconstruction, undistortion, and consistent projection between image and world coordinates. Why it matters for physical AI: Policies that consume RGB or depth images inherit any calibration error as systematic bias in perceived geometry, which is especially damaging for precise manipulation and for fusing multi-camera observations into a shared workspace frame. ### Camera Extrinsics Category: Perception URL: https://almond.bot/glossary/camera-extrinsics Camera extrinsics are the rigid-body transform, a rotation and translation forming an SE(3) pose, that relates the camera's coordinate frame to an external reference frame such as a robot base, end-effector, or world origin. Together with intrinsics, extrinsics allow 3D points to be projected into the image and pixels to be back-projected into workspace coordinates. Extrinsics are estimated via hand-eye calibration or fiducial markers and must be re-verified after any camera remounting. Why it matters for physical AI: Multi-view policies and 3D scene representations require accurate extrinsics to fuse observations; a few degrees of extrinsic error can move a predicted grasp point by centimeters, causing hard-to-diagnose policy failures. ### Camera Intrinsics Category: Perception URL: https://almond.bot/glossary/camera-intrinsics Camera intrinsics are the internal parameters of a camera's projection model: focal lengths, principal point, skew, and lens distortion coefficients, usually organized into a 3x3 intrinsic matrix K plus a distortion vector. They map 3D points in the camera frame to 2D pixel coordinates under the pinhole model. Intrinsics are fixed by the lens and sensor but change with zoom, focus, and temperature, and are estimated during camera calibration. Why it matters for physical AI: Depth back-projection, visual servoing, and any pixel-to-metric reasoning depend on correct intrinsics, and models trained on data from one camera can transfer poorly to cameras with different intrinsics unless images are rectified or normalized. ### Camera-to-Robot Calibration Category: Perception URL: https://almond.bot/glossary/camera-to-robot-calibration Camera-to-robot calibration is the estimation of the transform between a camera frame and a robot's kinematic frames, enabling perception outputs to be expressed in coordinates the robot can act on. Classical approaches solve the hand-eye equation AX = XB from paired robot and marker motions, while newer learning-based methods such as DREAM regress robot joint keypoints directly from RGB images, enabling markerless, self-supervised calibration. Both eye-in-hand and eye-to-hand configurations are common. Why it matters for physical AI: Every vision-guided action passes through this transform, so its accuracy bounds manipulation precision; markerless approaches also make large-scale, heterogeneous data collection practical by removing manual calibration from the loop. ### CAN Bus Category: Hardware URL: https://almond.bot/glossary/can-bus CAN Bus (Controller Area Network) is a robust serial communication protocol, originally developed by Bosch for automotive systems, that lets microcontrollers and devices exchange prioritized messages over a shared differential twisted pair without a host computer. In robotics it is widely used to command motor controllers and read encoder or sensor data, with classic CAN reaching 1 Mbps and CAN FD extending payloads and bitrates. Its arbitration scheme guarantees deterministic priority-based delivery. Why it matters for physical AI: Actuator command rates and latency over the bus set a hard floor on control frequency; understanding CAN bandwidth limits matters when scaling to high-DOF humanoids where dozens of joints share the same bus. ### Capstan Drive Category: Hardware URL: https://almond.bot/glossary/capstan-drive A capstan drive is a friction-based cable transmission in which a cable wraps around a driven drum and a larger output pulley, transmitting torque with effectively zero backlash and low friction. Because there are no meshing teeth, capstan drives offer smooth motion, high backdrivability, and quiet operation, at the cost of limited torque capacity and cable maintenance. They are used in haptic devices, some low-cost robot arms, and teleoperation interfaces where transparency matters. Why it matters for physical AI: Backdrivable, zero-backlash transmissions produce cleaner demonstration data during kinesthetic teaching and teleoperation, and make force-sensitive learned skills easier to execute than high-friction geared alternatives. ### Capture Point Category: Locomotion URL: https://almond.bot/glossary/capture-point The capture point is the location on the ground where a bipedal robot must step to bring itself to a complete stop, derived from the linear inverted pendulum model by Pratt et al. (2006). Formally the instantaneous capture point combines center-of-mass position and velocity scaled by the pendulum's natural frequency, and it generalizes to the divergent component of motion (DCM) used in modern humanoid controllers. Stepping controllers track it to recover from pushes. Why it matters for physical AI: Capture-point reasoning underlies push recovery and footstep placement in humanoid stacks, and remains a key model-based safety layer beneath learned locomotion policies deployed on real bipeds. ### Carbon Fiber Composite Category: Hardware URL: https://almond.bot/glossary/carbon-fiber-composite Carbon fiber composite is a material made of carbon fiber reinforcement embedded in a polymer matrix, offering exceptional stiffness-to-weight and strength-to-weight ratios compared to aluminum or steel. In robotics it is used for arm links, drone frames, and humanoid limb structures where reducing moving mass lowers inertia, boosts payload-to-weight ratio, and improves dynamic response. Drawbacks include cost, difficulty of machining, and brittle failure modes under impact. Why it matters for physical AI: Lighter links let actuators achieve faster, safer motions with less energy, directly improving the dynamic capabilities and battery life of mobile manipulators and humanoids operating around people. ### Cartesian Impedance Control Category: Control URL: https://almond.bot/glossary/cartesian-impedance-control Cartesian impedance control is a compliance control scheme that regulates the dynamic relationship between end-effector pose error and interaction force in task space, making the robot behave like a virtual spring-damper system attached to a target pose. Introduced conceptually by Hogan (1985), it maps Cartesian stiffness and damping through the manipulator Jacobian to joint torques, requiring torque-controlled joints. It enables safe, stable contact without explicit force setpoints. Why it matters for physical AI: Compliant task-space behavior lets learned policies command poses while the controller absorbs contact uncertainty, a common recipe for contact-rich skills like insertion and wiping on torque-controlled arms. ### Cartesian Robot Category: Hardware URL: https://almond.bot/glossary/cartesian-robot A Cartesian robot is a manipulator whose three principal axes are linear and mutually orthogonal, moving the end-effector along X, Y, and Z rails like a gantry. The kinematics are trivially decoupled, giving a rectangular workspace, high repeatability, and simple control, which is why the architecture dominates 3D printers, CNC machines, and pick-and-place gantries. Trade-offs include large footprint relative to workspace and limited dexterity compared to articulated arms. Why it matters for physical AI: Gantry-style robots remain the workhorses of structured automation, and their simple kinematics make them useful baselines and data collection platforms where full articulated dexterity is unnecessary. ### Cartesian Space (Task Space) Category: Math & Kinematics URL: https://almond.bot/glossary/cartesian-space Cartesian space, also called task space, is the coordinate space in which end-effector poses are described, typically position in R3 plus orientation, forming the special Euclidean group SE(3). It contrasts with joint (configuration) space, where states are joint angles. Motion goals for manipulation are naturally expressed in task space and mapped to joint space through inverse kinematics or differential kinematics via the Jacobian. Why it matters for physical AI: Most learned manipulation policies output task-space actions such as end-effector deltas because they transfer more readily across robots with different kinematics than raw joint commands. ### Cascade Control Category: Control URL: https://almond.bot/glossary/cascade-control Cascade control is a nested control architecture in which an outer loop's output serves as the setpoint for a faster inner loop, such as a position loop commanding a velocity loop that commands a current loop in a motor drive. The inner loop rejects fast disturbances locally, while the outer loop handles slower task-level objectives. Proper tuning requires the inner loop bandwidth to substantially exceed the outer loop's. Why it matters for physical AI: Nearly every robot actuator runs cascaded loops beneath learned policies; knowing which loop a policy commands, and at what rate, determines achievable smoothness, compliance, and the fidelity of executed actions. ### Causal Confusion Category: Robot Learning URL: https://almond.bot/glossary/causal-confusion Causal confusion is a failure mode of imitation learning in which a policy latches onto spurious correlates of expert actions rather than their true causes, formalized by de Haan et al. (2019). A classic example is a driving policy attending to its own brake indicator light instead of the pedestrian that caused braking. Counterintuitively, giving the policy more observation history or richer inputs can worsen the problem by adding confounders. Why it matters for physical AI: Behavior-cloned manipulation policies routinely exploit shortcuts like proprioceptive history or background cues, and diagnosing causal confusion guides choices about observation masking, history length, and data augmentation in large-scale training. ### Causal Inference (Robotics) Category: Robot Learning URL: https://almond.bot/glossary/causal-inference-robotics Causal inference in robotics is the use of causal models, interventions, and counterfactual reasoning to learn how a robot's actions affect the world, beyond the correlations captured by standard supervised learning. Because robots can act, they can perform interventions that identify cause-effect structure, distinguishing manipulable variables from confounded ones. Applications include disentangling action effects from environmental drift, transfer across domains, and diagnosing spurious features in learned policies. Why it matters for physical AI: Policies that capture causal rather than correlational structure generalize better under distribution shift, and interventional data from real robot fleets is a uniquely rich substrate for causal learning at scale. ### CE Marking Category: Safety & Standards URL: https://almond.bot/glossary/ce-marking CE Marking is the mandatory conformity mark indicating that a product placed on the European Economic Area market meets applicable EU health, safety, and environmental requirements. For robots this historically meant the Machinery Directive 2006/42/EC, being superseded by the Machinery Regulation (EU) 2023/1230, alongside EMC and, where relevant, Radio Equipment directives. Manufacturers perform risk assessment, compile a technical file, and issue a Declaration of Conformity, often demonstrating compliance via standards like ISO 10218. Why it matters for physical AI: Commercial deployment of robots in Europe is gated on CE conformity, and learned, adaptive behaviors raise novel questions in risk assessment that deployment teams must address in the technical file. ### Center of Mass Category: Math & Kinematics URL: https://almond.bot/glossary/center-of-mass The center of mass (CoM) is the mass-weighted average position of all particles in a body or multibody system, the point at which applied forces produce pure translation without rotation. For an articulated robot the CoM moves with configuration and is computed from link masses and forward kinematics. It is central to balance criteria such as the zero moment point and capture point, and to payload and stability analysis for mobile platforms. Why it matters for physical AI: Balance controllers for humanoids and mobile manipulators plan CoM trajectories explicitly, and learned locomotion policies implicitly regulate CoM dynamics, making accurate mass models important for sim-to-real transfer. ### Central Pattern Generator (CPG) Category: Locomotion URL: https://almond.bot/glossary/central-pattern-generator A central pattern generator is a neural circuit, biological or artificial, that produces rhythmic output without requiring rhythmic input, underlying gaits like walking and swimming in animals. In robotics, CPGs are implemented as coupled nonlinear oscillators whose phase relationships define gait patterns, providing smooth, stable rhythms modulated by simple high-level signals. Ijspeert's salamander robot work is a landmark demonstration, and CPG layers still appear beneath learned locomotion policies. Why it matters for physical AI: CPGs offer a compact, stable action space for legged locomotion learning: policies modulate oscillator parameters instead of raw torques, improving sample efficiency and sim-to-real robustness for gait control. ### Chain-of-Thought Planning Category: Foundation Models URL: https://almond.bot/glossary/chain-of-thought-planning Chain-of-thought planning is the use of intermediate reasoning steps, generated in natural language or structured text by a large language or vision-language model, to decompose a robot task before committing to actions. Rather than mapping observations directly to motor commands, the model reasons about subgoals, object relations, and preconditions, as explored in embodied reasoning systems and VLAs with reasoning traces. This improves long-horizon coherence and interpretability of robot behavior. Why it matters for physical AI: Explicit reasoning traces let robot foundation models handle multi-step tasks, expose auditable intermediate decisions for safety review, and transfer web-scale reasoning ability into embodied decision-making. ### Chunked Action Prediction Category: Robot Learning URL: https://almond.bot/glossary/chunked-action-prediction Chunked action prediction is a policy design in which the model outputs a sequence of future actions, an action chunk, per inference step rather than a single action, popularized by the ACT architecture in the ALOHA system (Zhao et al., 2023). Chunking reduces compounding error, mitigates causal confusion from idle demonstrator pauses, and lowers effective inference frequency requirements; temporal ensembling across overlapping chunks smooths execution. Diffusion Policy and most modern VLAs adopt chunked outputs. Why it matters for physical AI: Action chunking is a key enabler for running large policies on real robots, buying inference-time headroom while keeping motion smooth, and it strongly shapes the latency and reactivity trade-offs of deployed systems. ### Classifier-Free Guidance (CFG) Category: Robot Learning URL: https://almond.bot/glossary/classifier-free-guidance Classifier-free guidance is a sampling technique for conditional diffusion models, introduced by Ho and Salimans (2022), in which the model is trained with the conditioning signal randomly dropped, and at inference the conditional and unconditional score estimates are extrapolated to strengthen adherence to the condition. In robot learning it sharpens goal- or language-conditioning of diffusion policies, trading sample diversity for conditioning fidelity via a guidance weight. Why it matters for physical AI: Diffusion-based robot policies use guidance to make behavior follow language instructions or goals more reliably, and the guidance weight is a practical knob for tuning precision versus diversity at deployment. ### Clean Room Robot Category: Applications URL: https://almond.bot/glossary/clean-room-robot A clean room robot is a manipulator or mobile platform engineered to operate in contamination-controlled environments, classified under ISO 14644-1 cleanliness classes, common in semiconductor, pharmaceutical, and medical device manufacturing. Designs minimize particle generation using sealed joints, low-outgassing materials, special lubricants, and smooth stainless or coated surfaces, and are validated for particle emission. Wafer-handling robots in semiconductor fabs are the highest-volume example. Why it matters for physical AI: Semiconductor and pharma production are high-value automation domains where human presence itself is a contamination source, making cleanroom-rated autonomous manipulation an economically compelling deployment target. ### CLIP Category: Foundation Models URL: https://almond.bot/glossary/clip CLIP (Contrastive Language-Image Pre-training) is a vision-language model from OpenAI (Radford et al., 2021) that learns aligned image and text embeddings by contrastively training on 400 million web image-text pairs. Its dual encoders enable zero-shot classification, open-vocabulary retrieval, and semantic similarity scoring without task-specific fine-tuning. In robotics, CLIP embeddings ground language in perception for systems like CLIPort and open-vocabulary object detectors, and serve as reward or goal encoders. Why it matters for physical AI: CLIP-style aligned embeddings are the connective tissue between web-scale semantics and robot perception, enabling open-vocabulary grasping, language-specified goals, and semantic scene understanding without exhaustive robot-specific labels. ### Closed-Loop Control Category: Control URL: https://almond.bot/glossary/closed-loop-control Closed-loop control is a control strategy in which the system's output is measured and fed back to compute corrective commands, continuously driving the error between desired and actual state toward zero. Unlike open-loop execution, feedback rejects disturbances, model errors, and drift, at the cost of requiring sensing and raising stability considerations. PID, impedance, and model predictive controllers are all closed-loop, and the term also describes policies that re-observe before every action. Why it matters for physical AI: Reactive, closed-loop visuomotor policies distinguish modern robot learning from plan-then-execute pipelines: re-observing at every step lets policies correct slips, moved objects, and perception errors during execution. ### Cloth Manipulation Category: Manipulation URL: https://almond.bot/glossary/cloth-manipulation Cloth manipulation is the robotic handling of fabric and garments, spanning tasks like folding, flattening, unfolding, dressing assistance, and bed making. Cloth is a two-dimensional deformable object with near-infinite configuration space, severe self-occlusion, and complex contact dynamics, defying compact state representations. Research approaches include learned visual dynamics models, dense descriptor correspondence, flinging and other dynamic primitives, and large-scale imitation, with benchmarks such as SoftGym providing standardized evaluation. Why it matters for physical AI: Laundry folding and dressing are canonical household tasks that expose the gap between rigid-object competence and general manipulation, making cloth a key milestone for generalist home robot policies. ### Cloud Robotics Category: Software & Middleware URL: https://almond.bot/glossary/cloud-robotics Cloud robotics is an architecture in which robots offload computation, storage, and learning to remote data centers while retaining time-critical control onboard, a term popularized by James Kuffner around 2010. Typical patterns include cloud-hosted fleet management, shared maps and skill libraries, remote teleoperation gateways, and centralized training on fleet-collected data with periodic policy redeployment. Latency, bandwidth, and connectivity loss define the boundary between cloud and edge responsibilities. Why it matters for physical AI: Fleet learning loops, in which robots stream experience to central training clusters and receive improved policies, are the operational backbone of scaling robot foundation models with real-world data. ### Co-Training Category: Robot Learning URL: https://almond.bot/glossary/co-training Co-training in robot learning is the practice of training a single policy or model jointly on multiple data sources, such as data from different tasks, robots, simulators, or human video, so that shared structure improves each individual domain. Examples include RT-2's co-training on web vision-language data with robot trajectories and multi-robot co-training across embodiments in Open X-Embodiment. Careful data mixture weighting is typically decisive for final performance. Why it matters for physical AI: Robot data is scarce, so co-training with web-scale vision-language corpora and cross-robot datasets is currently the dominant strategy for injecting broad semantics and generalization into robot foundation models. ### Code as Policies Category: Foundation Models URL: https://almond.bot/glossary/code-as-policies Code as Policies is an approach from Liang et al. (Google, 2022) in which a large language model writes executable policy code, composing perception APIs, control primitives, and loops, directly from natural language instructions. Generated programs can invoke third-party libraries, define reusable functions, and express spatial reasoning numerically, enabling zero-shot generalization to novel instructions without policy training. It demonstrated that LLM code generation can serve as a robot planning interface. Why it matters for physical AI: Program synthesis gives language models a precise, verifiable interface to robot skills, complementing end-to-end VLAs with interpretable, editable behavior that engineers can inspect before execution. ### Collaborative Manufacturing Category: Applications URL: https://almond.bot/glossary/collaborative-manufacturing Collaborative manufacturing is a production paradigm in which humans and robots share workspaces and task sequences, dividing work according to complementary strengths: robots handle repetitive, heavy, or precise operations while humans contribute dexterity, judgment, and exception handling. It relies on collaborative robot safety modes such as speed and separation monitoring and power and force limiting, defined in ISO/TS 15066, plus workcell design that supports safe, fluent handovers. Why it matters for physical AI: Most near-term factory deployments place learned-skill robots beside human workers rather than replacing lines outright, so safe, fluent human-robot collaboration determines where physical AI actually enters production. ### Collaborative Robot (Cobot) Category: Hardware URL: https://almond.bot/glossary/collaborative-robot A collaborative robot, or cobot, is a manipulator designed to work in direct proximity to humans without conventional safety fencing, using power and force limiting, rounded geometry, and contact detection to keep potential injuries below thresholds defined in ISO/TS 15066. Popularized by Universal Robots' UR series from 2008, cobots typically feature hand guiding for programming, torque sensing in every joint, and payloads from 3 to 30 kg. Why it matters for physical AI: Cobots' torque sensing, backdrivability, and safe-contact design make them natural platforms for kinesthetic demonstration collection and for deploying learned skills in human-shared spaces without cage infrastructure. ### Collaborative Workspace Category: Safety & Standards URL: https://almond.bot/glossary/collaborative-workspace A collaborative workspace is the region within a robot cell where humans and robots may perform tasks simultaneously, as defined in ISO 10218 and detailed in ISO/TS 15066. Within it, one or more collaborative operating modes must apply: safety-rated monitored stop, hand guiding, speed and separation monitoring, or power and force limiting. Risk assessment determines the workspace boundaries, permitted speeds, and required sensing such as safety-rated laser scanners or vision. Why it matters for physical AI: Defining and monitoring shared space is the regulatory mechanism that lets autonomous manipulators operate around people, and perception-based dynamic safety zones are increasingly how learned systems meet these requirements. ### Collet Category: Hardware URL: https://almond.bot/glossary/collet A collet is a segmented, sleeve-like clamping device that grips cylindrical workpieces or tools by contracting uniformly around them when axially compressed, typically by a tapered nut or drawbar. Collets provide high concentricity and grip strength with fast changeover, making them standard in CNC spindles, lathes, and robotic tool changers. In robot end-of-arm tooling, collet-style mechanisms grip shafts, fasteners, and pegs where precise centering matters. Why it matters for physical AI: Precise, repeatable workpiece holding underpins machine-tending and assembly automation, and collet-based fixtures reduce the pose uncertainty that learned insertion and fastening policies must otherwise absorb. ### Collision Avoidance Category: Navigation & SLAM URL: https://almond.bot/glossary/collision-avoidance Collision avoidance is the capability of a robot to prevent unwanted contact with obstacles, people, and itself while executing motion, implemented through planning around known geometry, reactive methods like dynamic window approach or artificial potential fields, and control-level safety filters such as control barrier functions. It spans static and dynamic obstacles and requires reliable perception of the environment, whether from depth sensors, lidar, or learned occupancy predictions. Why it matters for physical AI: Learned policies do not inherently guarantee obstacle safety, so runtime avoidance layers and safety filters wrapped around neural policies are standard practice for deploying manipulation and navigation in unstructured spaces. ### Collision Checking Category: Navigation & SLAM URL: https://almond.bot/glossary/collision-checking Collision checking is the geometric computation that determines whether a robot configuration or motion segment intersects obstacles or the robot's own body, typically using libraries like FCL with broad-phase bounding-volume hierarchies and narrow-phase algorithms such as GJK. Sampling-based motion planners spend the majority of their runtime on collision checks, motivating GPU-parallel checkers such as cuRobo's and learned collision approximators. Continuous checking validates swept motions between sampled states. Why it matters for physical AI: Fast collision checking is the computational bottleneck of motion planning, and GPU-accelerated or learned checkers enable the real-time replanning rates that reactive manipulation around humans demands. ### Color Histogram Category: Perception URL: https://almond.bot/glossary/color-histogram A color histogram is a representation of the distribution of colors in an image or region, counting pixel frequencies across bins of a color space such as RGB or HSV. Histograms are invariant to translation and rotation and robust to partial occlusion, which made them a staple of classical object recognition, image retrieval, and tracking methods such as mean-shift and CAMShift. They discard all spatial structure, limiting discriminative power compared to learned features. Why it matters for physical AI: Though superseded by deep features for recognition, color statistics remain useful for lightweight tracking, dataset curation, and detecting domain shift between training and deployment imagery in robot vision pipelines. ### COLOSSEUM Category: Data & Benchmarks URL: https://almond.bot/glossary/colosseum COLOSSEUM is a simulation benchmark introduced in 2024 for evaluating the generalization of manipulation policies under systematic environmental perturbations. Built on RLBench tasks, it applies controlled variations across factors including object color, texture, size, lighting, distractor objects, camera pose, and physical properties, measuring how policy success degrades relative to the training conditions. It quantifies robustness gaps that standard same-distribution evaluations conceal. Why it matters for physical AI: Generalization under visual and physical variation is the central obstacle to deploying learned manipulation, and perturbation benchmarks like COLOSSEUM provide the controlled measurements needed to compare robustness across policy architectures. ### Compliance Category: Control URL: https://almond.bot/glossary/compliance Compliance is the ability of a robot to yield under external forces, the inverse of stiffness, arising passively from elastic mechanical elements or actively from control laws that modulate the relationship between force and displacement. Compliant behavior absorbs impacts, accommodates pose uncertainty during contact, and reduces peak forces in human-robot interaction. It is quantified by stiffness and damping parameters, often specified per Cartesian axis for manipulation tasks. Why it matters for physical AI: Contact-rich skills succeed on real hardware largely because compliance absorbs the residual errors of perception and learned policies, turning would-be jams and crashes into graceful accommodation. ### Compliant Actuator Category: Hardware URL: https://almond.bot/glossary/compliant-actuator A compliant actuator is an actuator that intentionally incorporates elasticity or low mechanical impedance between the motor and load, in contrast to stiff, highly geared designs. The canonical example is the series elastic actuator (Pratt and Williamson, 1995), which places a spring in the drivetrain to enable force sensing and shock tolerance; quasi-direct-drive and variable stiffness actuators are related approaches. Compliance improves safety and contact robustness at some cost in position bandwidth. Why it matters for physical AI: Actuator-level compliance and force transparency determine how safely robots can share space with people and how faithfully force-based demonstrations and learned contact skills execute on hardware. ### Compliant Grasp Category: Manipulation URL: https://almond.bot/glossary/compliant-grasp A compliant grasp is a grasp in which the gripper or hand yields elastically to conform to object geometry and accommodate pose uncertainty, achieved through soft materials, underactuated linkages, passive joints, or active force control. Mechanisms like the underactuated hands from the Harvard/Yale iRobot collaboration and soft grippers such as Festo's Fin Ray designs exploit compliance to grasp diverse objects without precise pre-grasp alignment. Why it matters for physical AI: Grasp-level compliance converts centimeter-scale perception and policy errors into successful grasps, dramatically lowering the precision burden on learned grasping systems handling diverse, unmodeled objects. ### Compliant Manipulation Category: Manipulation URL: https://almond.bot/glossary/compliant-manipulation Compliant manipulation is the execution of manipulation tasks while regulating interaction forces through mechanical or controlled compliance, rather than commanding pure position trajectories. Typical examples include peg-in-hole insertion, surface wiping and polishing, door opening, and assembly, where task geometry constrains motion and rigid position control would generate excessive forces. Implementations combine impedance or admittance control with strategies that exploit contact constraints to funnel uncertainty. Why it matters for physical AI: Most economically valuable manipulation involves sustained contact, and pairing learned policies with compliant control layers is the prevailing recipe for making such skills robust on real, imperfectly perceived hardware. ### Computed Torque Control Category: Control URL: https://almond.bot/glossary/computed-torque-control Computed torque control is a model-based control method that uses the robot's inverse dynamics to cancel nonlinear dynamics, applying feedback linearization so the closed-loop system behaves as a set of decoupled linear double integrators. Joint torques are computed from the mass matrix, Coriolis, and gravity terms evaluated along the desired trajectory, plus PD feedback on tracking error. Performance depends directly on the accuracy of the dynamic model. Why it matters for physical AI: Model-based torque control provides the precise, high-bandwidth tracking layer beneath many learned systems, and its sensitivity to model error motivates learned residual dynamics and adaptive extensions. ### Conditional Imitation Learning Category: Robot Learning URL: https://almond.bot/glossary/conditional-imitation-learning Conditional imitation learning is imitation learning in which the policy receives an additional conditioning input, such as a navigation command, goal image, or language instruction, that disambiguates which of several valid expert behaviors to reproduce. Introduced prominently for autonomous driving by Codevilla et al. (2018), where turn commands resolved intersection ambiguity, the idea generalizes to goal- and language-conditioned manipulation policies. Conditioning resolves the multimodality that plain behavior cloning averages away. Why it matters for physical AI: Generalist robot policies are inherently conditional, executing whichever task the instruction specifies, so conditioning architecture and grounding quality largely determine how well one model serves many tasks. ### Configuration Space (C-Space) Category: Math & Kinematics URL: https://almond.bot/glossary/configuration-space Configuration space is the set of all possible configurations of a robot, with each point specifying the values of every degree of freedom, such as the joint angles of a manipulator. Formalized for robotics by Lozano-Perez (1983), it transforms motion planning into finding a path for a point through the free subset of C-space, with obstacles mapped into C-space obstacles. Its dimension equals the robot's degrees of freedom, making high-DOF planning computationally demanding. Why it matters for physical AI: Sampling-based planners, collision checking, and the joint-space action representations of many learned policies all operate in configuration space, making it the mathematical substrate of motion generation. ### Conjugate Gradient Method Category: Math & Kinematics URL: https://almond.bot/glossary/conjugate-gradient-method The conjugate gradient method is an iterative algorithm for solving large systems of linear equations with symmetric positive-definite matrices, and by extension for unconstrained optimization, by searching along mutually conjugate directions. Developed by Hestenes and Stiefel (1952), it converges in at most n iterations for an n-dimensional exact problem and requires only matrix-vector products, making it memory efficient. Nonlinear variants extend it to general smooth objectives. Why it matters for physical AI: Conjugate gradient solvers appear inside physics engines for contact dynamics, in trajectory optimization, and in natural-gradient reinforcement learning methods like TRPO, quietly powering both simulation and policy optimization. ### Conservative Q-Learning (CQL) Category: Robot Learning URL: https://almond.bot/glossary/conservative-q-learning Conservative Q-Learning is an offline reinforcement learning algorithm from Kumar et al. (2020) that penalizes Q-values of actions outside the dataset distribution, learning a lower bound on the true value function to prevent the overestimation that plagues offline RL. The added regularizer pushes down Q-values for out-of-distribution actions while pushing up values for dataset actions, yielding policies that improve on the data without exploiting extrapolation errors. Why it matters for physical AI: Robots accumulate large logs of prior experience that cannot be safely extended by online trial and error; conservative offline RL methods extract improved policies from such fixed datasets without dangerous exploration. ### Consistency Policy Category: Robot Learning URL: https://almond.bot/glossary/consistency-policy Consistency Policy is a visuomotor policy approach (Prasad et al., 2024) that applies consistency distillation to diffusion policies, training a student network to map any point on the diffusion trajectory directly to its endpoint so actions are generated in one or a few inference steps. It retains the multimodal expressiveness of diffusion policies while cutting inference latency by an order of magnitude, enabling high-rate control on compute-constrained robots. Why it matters for physical AI: Diffusion policies' iterative sampling is often too slow for reactive control, so consistency-style distillation is a key path to running expressive generative policies at real-time rates on embedded hardware. ### Constraint Learning Category: Robot Learning URL: https://almond.bot/glossary/constraint-learning Constraint learning is the inference of task or safety constraints, such as regions to avoid, forces not to exceed, or invariants to maintain, from demonstrations, corrections, or environment interaction, rather than specifying them manually. Approaches include inverse constraint learning from demonstrations assumed to be constraint-satisfying, learning constraint manifolds for planning, and extracting geometric task constraints like axis alignment from a handful of examples. Why it matters for physical AI: Explicit learned constraints complement end-to-end policies by encoding hard requirements that must hold even off-distribution, supporting safety filters and verifiable behavior in deployed systems. ### Construction Robot Category: Applications URL: https://almond.bot/glossary/construction-robot A construction robot is a robotic system that performs building tasks such as bricklaying, welding, drilling, layout marking, rebar tying, 3D-printing structures, or site surveying. Examples include semi-autonomous layout robots, drone-based site scanning, and quadrupeds like Boston Dynamics' Spot used for progress monitoring. Construction combines unstructured outdoor environments, heavy materials, and tight tolerances, making it one of the hardest and least automated major industries. Why it matters for physical AI: Construction's labor shortages and low productivity growth create strong demand for autonomy, while its unstructured, constantly changing sites are a proving ground for robust perception and mobile manipulation. ### Contact Force Estimation Category: Control URL: https://almond.bot/glossary/contact-force-estimation Contact force estimation is the inference of external forces and torques acting on a robot, either from dedicated force-torque sensors or indirectly from joint torque measurements, motor currents, and dynamic models via momentum-observer or disturbance-observer techniques. Sensorless estimation enables collision detection and force control on arms without wrist sensors, though accuracy is limited by model quality and friction. Learned estimators increasingly fuse proprioception, currents, and vision. Why it matters for physical AI: Force awareness without expensive sensing on every joint enables safe contact detection and force-conditioned policies at fleet scale, and estimated wrenches are valuable labels for training contact-rich skills. ### Contact Mode Category: Manipulation URL: https://almond.bot/glossary/contact-mode A contact mode is a discrete classification of the contact state between bodies, specifying which contacts are active and whether each is sticking, sliding, or separating. Manipulation tasks traverse sequences of contact modes, and each mode imposes different dynamics and constraints, making the overall system hybrid. Contact-implicit trajectory optimization and mode-enumeration planners reason over these discrete switches, whose combinatorial explosion is a core difficulty of dexterous manipulation. Why it matters for physical AI: Learned policies implicitly discover contact mode sequences that model-based planners must enumerate explicitly, and understanding mode structure explains why contact-rich tasks are so much harder than free-space motion. ### Contact Schedule Category: Locomotion URL: https://almond.bot/glossary/contact-schedule A contact schedule is the planned timing and sequence of contact events, specifying which feet or bodies touch which surfaces over a motion, used extensively in legged locomotion and increasingly in loco-manipulation. Given a schedule, the continuous trajectory optimization becomes smooth within each phase, so many controllers fix gait timings a priori, while contact-implicit optimization discovers schedules automatically at greater computational cost. Why it matters for physical AI: Whether contact timing is fixed, optimized, or learned end-to-end is a central design axis in legged robot controllers, shaping agility, robustness to slips, and the difficulty of the learning problem. ### Contact Simulation Category: Simulation URL: https://almond.bot/glossary/contact-simulation Contact simulation is the numerical modeling of forces and motion arising when rigid or deformable bodies touch, requiring collision detection, contact force resolution, and friction modeling, typically via linear complementarity formulations, penalty methods, or MuJoCo's smooth convex approximation. Accurate, stable, and fast contact resolution is the hardest part of robotics simulation, and discrepancies between simulated and real contact dynamics are a principal source of the sim-to-real gap. Why it matters for physical AI: Policies for manipulation and locomotion are predominantly trained in simulation, so contact model fidelity and speed directly bound what skills can be learned and how well they transfer to hardware. ### Contact-Rich Manipulation Category: Manipulation URL: https://almond.bot/glossary/contact-rich-manipulation Contact-rich manipulation is manipulation in which sustained, force-significant contact between the robot, object, and environment is essential to the task, as in insertion, assembly, wiping, prying, and in-hand reorientation. Such tasks feature hybrid dynamics, friction uncertainty, and partial observability of contact state, defeating pure position control. Successful systems combine compliance, force feedback, tactile sensing, or policies trained on real interaction data to exploit rather than avoid contact. Why it matters for physical AI: Industrial assembly, logistics, and household work are dominated by contact-rich steps, making them the decisive test of whether learned policies deliver economic value beyond pick-and-place. ### Context-Conditioned Policy Category: Robot Learning URL: https://almond.bot/glossary/context-conditioned-policy A context-conditioned policy is a policy whose behavior is modulated by a context variable summarizing task, environment, or embodiment properties, such as a latent vector inferred from recent experience, a task embedding, or explicit system parameters. In meta-RL methods like PEARL, the context is inferred online from transitions, enabling rapid adaptation; in multi-task learning, contexts index skills. The policy shares parameters across contexts, amortizing learning. Why it matters for physical AI: Fleet-scale deployment demands single models that adapt to varied robots, payloads, and sites; context conditioning provides the mechanism for such adaptation without per-site retraining. ### Continual Learning Category: Robot Learning URL: https://almond.bot/glossary/continual-learning Continual learning is the ability of a model to acquire new tasks or adapt to new conditions sequentially without catastrophically forgetting previously learned capabilities. Techniques include regularization methods like elastic weight consolidation, replay of stored or generated past experience, and parameter isolation via adapters or expansion. For robots, continual learning spans new skills, new environments, and hardware drift, ideally from the robot's own ongoing deployment experience. Why it matters for physical AI: Deployed robots encounter novelty daily, and updating fleet policies with fresh experience while preserving validated skills is essential for improving foundation models after they leave the lab. ### Continuous Control Category: Robot Learning URL: https://almond.bot/glossary/continuous-control Continuous control is the class of control and reinforcement learning problems whose action spaces are continuous vectors, such as joint torques, velocities, or end-effector twists, rather than discrete choices. It demands policy representations and RL algorithms that handle infinite action sets, driving the development of methods like DDPG, TD3, SAC, and PPO, commonly benchmarked on MuJoCo locomotion suites and DeepMind Control. Why it matters for physical AI: Virtually all robot motor commands are continuous, so advances in continuous-action policy learning translate directly into what robots can physically execute, from locomotion gaits to dexterous finger coordination. ### Contour Detection Category: Perception URL: https://almond.bot/glossary/contour-detection Contour detection is the identification of object boundaries and salient edges in images, producing curves that separate regions, classically via edge detectors like Canny followed by contour tracing, and in modern pipelines via learned boundary networks or the boundaries of predicted segmentation masks. Contours support shape analysis, pose estimation of known parts, bin picking of rigid components, and visual servoing on object silhouettes. Why it matters for physical AI: Boundary cues remain valuable for precise part localization in industrial vision and for extracting grasp-relevant geometry when full 3D perception is unavailable or too slow. ### Contraction Theory Category: Control URL: https://almond.bot/glossary/contraction-theory Contraction theory is a framework for analyzing nonlinear system stability by studying whether trajectories converge toward each other exponentially, rather than toward a fixed equilibrium, formalized by Lohmiller and Slotine (1998) through the differential dynamics and a contraction metric. Contracting systems forget initial conditions and compose well under series, parallel, and feedback combinations, enabling modular stability certificates for observers, controllers, and learned dynamics. Why it matters for physical AI: Contraction metrics provide tractable stability certificates for learned dynamics models and neural controllers, supporting certifiable learning-based control where trajectory-convergence guarantees are required for deployment. ### Contrastive Learning Category: Robot Learning URL: https://almond.bot/glossary/contrastive-learning Contrastive learning is a self-supervised representation learning approach that trains encoders to pull embeddings of positive pairs, different views or modalities of the same underlying content, together while pushing apart negatives, typically using the InfoNCE loss. Landmark methods include SimCLR and MoCo for images and CLIP for image-text alignment. In robotics, contrastive objectives learn visual representations from unlabeled robot video, align observations with goals, and pretrain encoders for policies. Why it matters for physical AI: Contrastive pretraining converts abundant unlabeled robot and human video into useful visual representations, cutting the demonstration count needed to train manipulation policies and improving robustness to visual variation. ### Contrastive Predictive Coding (CPC) Category: Robot Learning URL: https://almond.bot/glossary/contrastive-predictive-coding Contrastive Predictive Coding is a self-supervised learning method from van den Oord et al. (2018) that learns representations by predicting future latent states in a sequence and distinguishing true futures from negatives via the InfoNCE loss, which maximizes a lower bound on mutual information. Applied to audio, images, text, and RL, CPC-style temporal contrast underlies representation learning from sequential robot experience and video. Why it matters for physical AI: Temporal predictive objectives extract dynamics-aware features from raw robot experience without labels, providing pretrained encoders that accelerate downstream policy and world-model learning. ### Control Barrier Function (CBF) Category: Control URL: https://almond.bot/glossary/control-barrier-function A control barrier function is a scalar function of the system state whose sublevel or superlevel set defines a safe region, together with a condition on its derivative ensuring the set is forward invariant under control, formalized by Ames et al. CBF constraints are typically enforced pointwise via quadratic programs that minimally modify a nominal controller's command, yielding real-time safety filters for collision avoidance, joint limits, and force bounds. Why it matters for physical AI: CBF-based safety filters wrap learned policies with mathematically grounded runtime guarantees, letting neural controllers act freely inside a certified safe set while overriding them near its boundary. ### Control Lyapunov Function (CLF) Category: Control URL: https://almond.bot/glossary/control-lyapunov-function A control Lyapunov function is a positive-definite, energy-like function of the state for which some admissible control input can always make its time derivative negative, certifying that the system can be driven toward an equilibrium or goal set. Introduced by Artstein and Sontag in the 1980s, CLFs turn stabilization into a pointwise constraint solvable by quadratic programming, and are frequently combined with control barrier functions to unify stability and safety. Why it matters for physical AI: CLF certificates provide provable convergence for goal-reaching behaviors, and learning neural Lyapunov functions for complex dynamics is an active route to verified stability in learned robot controllers. ### Convolutional Neural Network (CNN) Category: Perception URL: https://almond.bot/glossary/convolutional-neural-network A convolutional neural network is a neural architecture that applies learned convolutional filters with shared weights across spatial positions, building translation-equivariant hierarchies of visual features. From LeNet through AlexNet (2012) and ResNet (2015), CNNs drove the deep learning revolution in computer vision and remain standard visual encoders in robot policies, including the ResNet backbones of many behavior cloning and diffusion policy implementations, though vision transformers now rival them. Why it matters for physical AI: CNN encoders remain a compute-efficient choice for real-time visuomotor policies on embedded hardware, and their inductive biases help when robot datasets are too small to train transformers from scratch. ### Cooling System Category: Hardware URL: https://almond.bot/glossary/cooling-system A cooling system is the thermal management hardware that removes heat from a robot's motors, drives, and compute, using passive heatsinks, forced-air fans, or liquid cooling loops. Actuator torque is thermally limited: continuous torque ratings assume steady-state heat dissipation, and exceeding them triggers derating or shutdown. High-torque-density humanoid actuators and embedded GPUs both push thermal envelopes, making cooling design integral to sustained performance. Why it matters for physical AI: Sustained manipulation workloads and onboard inference are thermally bound, so cooling capacity determines whether a robot's peak capabilities are usable for minutes or full shifts in deployment. ### Coriolis Forces Category: Math & Kinematics URL: https://almond.bot/glossary/coriolis-forces Coriolis forces are velocity-dependent inertial forces arising in robot dynamics from the coupling between joint velocities, appearing alongside centrifugal terms in the manipulator equation of motion as the Coriolis matrix multiplied by joint velocities. They grow quadratically with speed and couple motion across joints, so fast trajectories require their compensation in model-based controllers such as computed torque control. They vanish at zero velocity, which is why slow motions tolerate simpler control. Why it matters for physical AI: Dynamic, high-speed skills like throwing and fast pick cycles demand controllers or learned models that account for velocity-coupled dynamics, a key sim-to-real consideration for aggressive learned motions. ### Correspondence Problem Category: Perception URL: https://almond.bot/glossary/correspondence-problem The correspondence problem is the task of identifying which elements in one image, point cloud, or observation match the same physical entity in another, underlying stereo depth, optical flow, feature-based SLAM, point cloud registration, and cross-instance semantic matching. Solutions range from handcrafted descriptors like SIFT to learned dense descriptors such as Dense Object Nets, which enable category-level manipulation by matching keypoints across object instances. Why it matters for physical AI: Dense learned correspondences let robots transfer grasps and skills across object instances and viewpoints, providing spatial grounding that complements the semantic knowledge of vision-language models. ### Cosine Annealing Category: Robot Learning URL: https://almond.bot/glossary/cosine-annealing Cosine annealing is a learning rate schedule that decays the learning rate from its initial value to a minimum following a half-cosine curve over training, introduced by Loshchilov and Hutter (2017) as part of SGDR, which also proposed warm restarts. The smooth decay avoids abrupt transitions of step schedules, and combined with linear warmup it is the default schedule for training transformers, including vision-language-action models. Why it matters for physical AI: Training recipes for large robot policies are sensitive to schedule choices, and cosine decay with warmup is the de facto standard that practitioners tune when pretraining or fine-tuning foundation models on robot data. ### Cost Function Category: Control URL: https://almond.bot/glossary/cost-function A cost function is a scalar objective that quantifies the undesirability of states, actions, or trajectories, which optimization-based planners and controllers minimize, the mirror image of a reward function being maximized. In optimal control, costs typically penalize state error, control effort, and constraint proximity, as in the quadratic costs of LQR and MPC; in motion planning, costs encode path length, smoothness, and obstacle clearance. Why it matters for physical AI: Cost design encodes intent for model-based layers just as reward design does for RL, and mis-specified costs produce confidently wrong behavior, motivating learned costs from demonstrations via inverse optimal control. ### Costmap Category: Navigation & SLAM URL: https://almond.bot/glossary/costmap A costmap is a grid-based representation of the environment in which each cell stores a traversal cost, typically built from occupancy data with obstacles marked lethal and surrounding cells inflated by the robot's footprint and a safety margin. Layered costmaps, as in ROS navigation stacks like Nav2, compose static maps, live sensor obstacles, and semantic or social layers. Planners search the costmap for minimum-cost, collision-free paths. Why it matters for physical AI: Costmaps remain the practical interface between perception and motion for mobile robots, and learned semantic layers, such as terrain traversability predicted by neural networks, plug naturally into this representation. ### Covariance Matrix Category: Math & Kinematics URL: https://almond.bot/glossary/covariance-matrix A covariance matrix is a symmetric positive semi-definite matrix whose entries are the pairwise covariances of a random vector's components, characterizing the spread and correlation of uncertainty around the mean. In robotics it represents state uncertainty in Kalman filters and SLAM, sensor noise models, and pose estimation confidence, with its eigenstructure describing uncertainty ellipsoids. Propagating covariance through motion and measurement models is the core computation of Gaussian state estimation. Why it matters for physical AI: Calibrated uncertainty lets robots know what they do not know, informing when to re-observe, slow down, or defer to a human, which is fundamental to safe autonomy under imperfect perception. ### Coverage Path Planning Category: Navigation & SLAM URL: https://almond.bot/glossary/coverage-path-planning Coverage path planning is the computation of a path that passes over all points of a target area or surface, subject to the robot's footprint or tool width, used in floor cleaning, lawn mowing, agricultural field operations, spray painting, and inspection. Classical methods include boustrophedon (back-and-forth) cellular decomposition and spiral patterns, with extensions handling obstacles, energy limits, and multi-robot partitioning. Why it matters for physical AI: Coverage tasks constitute some of the largest deployed robot fleets, from household vacuums to agricultural machinery, and pairing coverage planners with learned terrain and dirt perception improves their real-world efficiency. ### Crop Harvesting Category: Applications URL: https://almond.bot/glossary/crop-harvesting Crop harvesting robotics is the automation of picking fruits, vegetables, and other crops, combining detection of ripe produce under occlusion and variable lighting, motion planning through foliage, and gentle detachment with soft or specialized grippers. Systems exist for strawberries, apples, tomatoes, and sweet peppers, though commercial viability hinges on picking speed and damage rates relative to human crews. Seasonal labor shortages drive sustained investment. Why it matters for physical AI: Harvesting stresses nearly every hard problem at once: cluttered outdoor perception, delicate force control, and economic throughput targets, making agriculture a demanding proving ground for general manipulation. ### Cross Product Category: Math & Kinematics URL: https://almond.bot/glossary/cross-product The cross product is a binary operation on two vectors in R3 producing a vector perpendicular to both, with magnitude equal to the parallelogram area they span and direction given by the right-hand rule. In robotics it computes torque as r cross F, angular momentum, and velocity of rotating points, and its skew-symmetric matrix form underlies the Jacobian columns of revolute joints and the algebra of rotations in se(3). Why it matters for physical AI: Torque computation, Jacobian construction, and rigid-body dynamics all reduce to cross products, so the operation appears in virtually every kinematics and dynamics routine executing on a robot. ### Cross-Attention Category: Foundation Models URL: https://almond.bot/glossary/cross-attention Cross-attention is the transformer mechanism in which queries from one sequence attend over keys and values from another, fusing information across modalities or streams, in contrast to self-attention within a single sequence. Introduced with the transformer's encoder-decoder attention (Vaswani et al., 2017), it is how vision-language-action models condition action generation on visual tokens and language instructions, and how diffusion policies inject observation embeddings into denoising networks. Why it matters for physical AI: Cross-attention is the primary architectural fusion point in modern robot policies, binding language goals, multi-view images, and proprioception into unified action prediction; its design shapes grounding quality and inference cost. ### Cross-Embodiment Transfer Category: Robot Learning URL: https://almond.bot/glossary/cross-embodiment-transfer Cross-embodiment transfer is the ability of a policy or model trained on data from one or more robot embodiments to work on, or accelerate learning for, robots with different kinematics, sensors, or action spaces. The Open X-Embodiment project (2023) pooled data from over 20 robot types to show that co-trained RT-X models outperform single-robot training, and models like CrossFormer extend this across manipulation, navigation, and locomotion. Why it matters for physical AI: No single platform will generate enough data alone, so transferring skills across heterogeneous fleets is a load-bearing hypothesis behind scaling robot foundation models to general competence. ### Cross-Entropy Loss Category: Robot Learning URL: https://almond.bot/glossary/cross-entropy-loss Cross-entropy loss is a loss function measuring the dissimilarity between a predicted probability distribution and target labels, equal to the negative log-likelihood of the correct class under the model. It is the standard objective for classification and for next-token prediction in language models, and appears in robot learning when policies output discretized action bins, as in RT-1 and RT-2, or when VLAs are trained autoregressively on action tokens. Why it matters for physical AI: Tokenized-action VLAs are trained almost entirely with cross-entropy, so its properties, including how action discretization interacts with likelihood training, directly shape the precision of foundation-model robot control. ### Cross-Roller Bearing Category: Hardware URL: https://almond.bot/glossary/cross-roller-bearing A cross-roller bearing is a compact rotary bearing in which cylindrical rollers are arranged alternately at 90 degrees to each other in a single V-groove raceway, allowing one bearing to support radial, axial, and moment loads simultaneously with high rigidity. This makes them standard as the output bearing of robot joints, integrated into harmonic drive units, and in rotary tables, where they resist tilting moments from cantilevered arm loads. Why it matters for physical AI: Joint stiffness and moment capacity set achievable end-effector precision under load, so output bearing quality quietly bounds the repeatability that learned fine-manipulation skills can exploit. ### CrossFormer Category: Foundation Models URL: https://almond.bot/glossary/crossformer CrossFormer is a cross-embodiment robot policy from UC Berkeley (Doshi et al., 2024), a transformer trained on 900K trajectories spanning single-arm and bimanual manipulation, quadruped locomotion, and navigation without requiring a shared observation or action space. Variable observation tokens and per-embodiment action readouts let one architecture control drastically different robots, matching specialist policies while enabling transfer across embodiments. Why it matters for physical AI: CrossFormer demonstrated that one policy can span manipulation, navigation, and locomotion without aligning action spaces, strengthening the case for unified generalist models over per-platform specialists. ### Curiosity-Driven Exploration Category: Robot Learning URL: https://almond.bot/glossary/curiosity-driven-exploration Curiosity-driven exploration is a reinforcement learning technique that supplements or replaces external rewards with intrinsic rewards for encountering novel or poorly predicted states, encouraging systematic exploration in sparse-reward settings. Notable formulations include prediction-error curiosity via the Intrinsic Curiosity Module (Pathak et al., 2017) and Random Network Distillation. Care is required to avoid pathologies like the noisy-TV problem, where irreducible stochasticity attracts endless attention. Why it matters for physical AI: Real-world tasks rarely come with dense rewards, and intrinsic motivation offers robots a way to autonomously gather diverse, informative experience for training skills and world models without hand-crafted supervision. ### Curriculum Learning Category: Robot Learning URL: https://almond.bot/glossary/curriculum-learning Curriculum learning is a training strategy that presents tasks or data in a progression from easy to hard, formalized by Bengio et al. (2009), so that competence on simpler instances scaffolds learning of harder ones. In robot RL, curricula schedule terrain difficulty, goal distance, randomization ranges, or disturbance magnitudes, with automatic variants adapting difficulty to current performance; the technique was central to learned quadruped locomotion over challenging terrain. Why it matters for physical AI: Curricula frequently make the difference between success and failure in sim-based skill learning, enabling policies for parkour, rough terrain, and dexterous tasks that direct training on full difficulty never reaches. ### Cycle Time Category: Applications URL: https://almond.bot/glossary/cycle-time Cycle time is the elapsed duration for a robot or workcell to complete one full unit of its repetitive task, such as one pick-and-place or one assembly operation, directly determining throughput. Industrial automation is specified and purchased against cycle time targets, and optimization involves trajectory time-optimality, motion blending, gripper actuation overlap, and minimizing sensing dwell. It is typically reported alongside success rate. Why it matters for physical AI: Learned policies must compete with hard-coded automation on throughput, not just flexibility, so inference latency and motion efficiency translate directly into the cycle times that decide commercial adoption. ### Cycloidal Drive Category: Hardware URL: https://almond.bot/glossary/cycloidal-drive A cycloidal drive is a compact, high-ratio gear reducer in which an eccentric input drives a cycloidal disc that rolls against ring pins, producing output rotation with high shock-load tolerance, high torsional stiffness, and low backlash. Compared to harmonic drives, cycloidal reducers handle higher impact loads and offer greater rigidity at the cost of more mass and slight torque ripple, making them common in the proximal joints of industrial robots. Why it matters for physical AI: Reducer choice shapes a joint's stiffness, backlash, efficiency, and impact survivability, all of which condition the dynamics that learned controllers must master and the abuse hardware tolerates in deployment. ## Glossary: D ### D* Lite Category: Navigation & SLAM URL: https://almond.bot/glossary/d-star-lite D* Lite is an incremental heuristic search algorithm for path planning in unknown or changing environments, introduced by Koenig and Likhachev in 2002 as a simpler reformulation of Stentz's D*. Built on Lifelong Planning A*, it efficiently repairs a previously computed shortest path when edge costs change instead of replanning from scratch. Algorithms in the D* family have been used for rover navigation, including planetary exploration systems. Why it matters for physical AI: Mobile robots and manipulator bases constantly discover obstacles their maps missed. Incremental replanners like D* Lite keep navigation reactive at real-world update rates, a prerequisite for deploying autonomous platforms outside static, fully mapped facilities. ### DAgger (Dataset Aggregation) Category: Robot Learning URL: https://almond.bot/glossary/dagger DAgger (Dataset Aggregation) is an iterative imitation learning algorithm introduced by Ross, Gordon, and Bagnell in 2011. The learner executes its current policy, an expert relabels the visited states with correct actions, and the aggregated dataset is used to retrain the policy. This directly addresses the covariate shift problem of behavior cloning, in which small errors compound because the policy encounters states absent from the expert's demonstrations. Why it matters for physical AI: Compounding errors are a dominant failure mode of demonstration-trained robot policies. DAgger-style corrective labeling, including human-gated variants used in real teleoperation pipelines, remains a practical recipe for closing the gap between offline training and closed-loop execution. ### Data Augmentation Category: Robot Learning URL: https://almond.bot/glossary/data-augmentation Data Augmentation is the practice of expanding a training dataset by applying label-preserving transformations to existing samples. In robot learning this spans image-level perturbations such as color jitter, random cropping, and noise injection, as well as robotics-specific strategies like mirroring bimanual trajectories, replaying demonstrations from novel simulated viewpoints, and randomizing distractor objects. Random crop augmentation was a key ingredient in sample-efficient visual RL methods such as DrQ. Why it matters for physical AI: Real robot data is expensive to collect, so squeezing more generalization out of every demonstration is critical. Well-chosen augmentations improve robustness to lighting, camera pose, and scene variation without additional hardware time. ### Data Distribution Service (DDS) Category: Software & Middleware URL: https://almond.bot/glossary/data-distribution-service Data Distribution Service (DDS) is an Object Management Group middleware standard for real-time, data-centric publish-subscribe communication. It provides decentralized peer discovery and fine-grained quality-of-service policies covering reliability, durability, and deadlines. DDS is the default transport layer underlying ROS 2, with interchangeable vendor implementations such as Fast DDS and Cyclone DDS. Why it matters for physical AI: Robot software stacks must move sensor and control messages between many processes with predictable latency. DDS's quality-of-service guarantees and brokerless architecture underpin the reliability of ROS 2 systems deployed on real hardware. ### Data Flywheel Category: Data & Benchmarks URL: https://almond.bot/glossary/data-flywheel Data Flywheel is a self-reinforcing loop in which deployed systems generate data that improves the underlying models, and the improved models drive broader deployment that generates still more data. The concept was popularized by autonomous driving fleets and is now central to robot foundation model strategy, where fleet telemetry, interventions, and successful rollouts are fed back into training. Why it matters for physical AI: Robotics lacks an internet-scale corpus of action-labeled data, so companies that close the deployment-to-training loop can compound an advantage. Fleet learning turns every commissioned robot into a data collection asset. ### Data Loader Category: Software & Middleware URL: https://almond.bot/glossary/data-loader Data Loader is the component of a training pipeline that reads samples from storage, decodes and transforms them, and assembles shuffled batches for the optimizer. Robot learning loaders must handle multimodal, temporally aligned streams of images, proprioception, and actions, often sampling windows from long episodes stored in formats such as RLDS or the LeRobot dataset format. Throughput matters: video decoding frequently bottlenecks policy training more than the forward pass. Why it matters for physical AI: Training visuomotor policies on hundreds of hours of multi-camera video is an infrastructure problem as much as a modeling one. Efficient loading and chunk sampling directly determine how fast large robot datasets can be iterated on. ### Data Scaling Laws Category: Robot Learning URL: https://almond.bot/glossary/data-scaling-laws Data Scaling Laws are empirical relationships describing how model performance improves predictably, typically as a power law, with increases in training data, parameters, or compute. Established for language models by Kaplan et al. (2020) and refined by the Chinchilla analysis, analogous studies in robotics examine how policy generalization scales with the number of demonstrations, environments, and embodiments in the training mix. Why it matters for physical AI: Scaling laws let teams forecast how much data collection is needed to hit a target success rate before spending months on it. Early robotics results suggest environment and task diversity matter as much as raw demonstration count. ### Dataset Curation Category: Data & Benchmarks URL: https://almond.bot/glossary/dataset-curation Dataset Curation is the process of selecting, filtering, deduplicating, and rebalancing training data to improve downstream model quality. For robot datasets this includes removing failed or idle segments, scoring demonstration quality and operator consistency, balancing task and scene distributions, and verifying calibration and timestamp alignment across sensor streams. Curation choices for large corpora such as Open X-Embodiment materially affect what policies trained on them learn. Why it matters for physical AI: Noisy or suboptimal demonstrations are cloned along with good ones, so curation quality often bounds imitation learning performance. Systematic filtering and weighting can outperform simply collecting more raw data. ### Dead Reckoning Category: Navigation & SLAM URL: https://almond.bot/glossary/dead-reckoning Dead Reckoning is the estimation of a robot's current pose by integrating motion measurements, such as wheel odometry, IMU readings, or commanded velocities, from a known starting point. Because each measurement carries noise and bias, integration causes the pose estimate to drift without bound over time. Practical systems therefore fuse dead reckoning with absolute corrections from GPS, visual landmarks, or map matching. Why it matters for physical AI: Drift-prone odometry is the default state estimate whenever exteroceptive sensing drops out, so its error characteristics shape sensor fusion design. Understanding dead reckoning failure modes is essential for robust localization in warehouses, homes, and outdoor deployments. ### Decision Transformer Category: Robot Learning URL: https://almond.bot/glossary/decision-transformer Decision Transformer is an offline reinforcement learning approach, introduced by Chen et al. in 2021, that casts control as conditional sequence modeling. A causal transformer is trained on trajectories of returns-to-go, states, and actions, and at test time is conditioned on a desired return to autoregressively generate actions. This reframing replaces value bootstrapping and policy gradients with supervised sequence prediction. Why it matters for physical AI: Treating control as sequence modeling lets robotics borrow the architectures, tooling, and scaling behavior of large language models. The paradigm influenced action-as-token designs now common in vision-language-action models. ### Deep Deterministic Policy Gradient (DDPG) Category: Robot Learning URL: https://almond.bot/glossary/deep-deterministic-policy-gradient Deep Deterministic Policy Gradient (DDPG) is an off-policy actor-critic algorithm for continuous action spaces, introduced by Lillicrap et al. in 2015. It combines the deterministic policy gradient theorem with DQN-style innovations, using a replay buffer and target networks to stabilize learning of a Q-function critic and a deterministic actor. Successors such as TD3 and SAC address its overestimation bias and brittleness to hyperparameters. Why it matters for physical AI: Continuous torque and velocity commands are the native action space of robots, and DDPG was among the first deep RL methods to handle them directly. Its descendants still power simulation-trained locomotion and manipulation controllers. ### Deep Q-Network (DQN) Category: Robot Learning URL: https://almond.bot/glossary/deep-q-network Deep Q-Network (DQN) is a reinforcement learning algorithm that approximates the action-value function with a deep neural network, introduced by Mnih et al. and shown in 2015 to reach human-level play on Atari games from raw pixels. Its key stabilizing mechanisms, experience replay and a periodically updated target network, became standard components of value-based deep RL. DQN operates over discrete action sets. Why it matters for physical AI: DQN demonstrated that control policies could be learned end-to-end from pixels, a founding result for learned visuomotor control. Its replay and target-network machinery persists in the off-policy algorithms used to train robot skills in simulation. ### Deep Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/deep-reinforcement-learning Deep Reinforcement Learning is the combination of reinforcement learning with deep neural network function approximators for policies, value functions, or world models. Landmark results include DQN on Atari, AlphaGo, and OpenAI's dexterous in-hand manipulation. In robotics it is most commonly applied in massively parallel simulation, with algorithms such as PPO and SAC, followed by transfer of the trained policy to hardware. Why it matters for physical AI: Simulation-scale deep RL produces controllers, particularly for legged locomotion and dexterous manipulation, that are difficult to hand-engineer or demonstrate. Combined with domain randomization it is a primary route from GPU-hours to deployable low-level skills. ### Deformable Object Manipulation Category: Manipulation URL: https://almond.bot/glossary/deformable-object-manipulation Deformable Object Manipulation is the robotic handling of objects that change shape under contact, including cloth, cables, bags, and food items. It is challenging because the object state is effectively infinite-dimensional, self-occlusion complicates perception, and dynamics are difficult to model or simulate accurately. Benchmarks such as SoftGym and tasks like cloth folding, rope routing, and bag opening drive research combining learned dynamics models, dense visual correspondence, and imitation learning. Why it matters for physical AI: A large share of household, hospital, and logistics work involves laundry, cables, packaging, and food rather than rigid parts. Progress on deformables is therefore a gating factor for general-purpose robots leaving structured industrial settings. ### Degrees of Freedom (DOF) Category: Math & Kinematics URL: https://almond.bot/glossary/degrees-of-freedom Degrees of Freedom (DOF) are the number of independent parameters required to fully specify the configuration of a mechanical system. A rigid body in space has six degrees of freedom, three translational and three rotational, so a manipulator needs at least six joints to reach arbitrary end-effector poses; a seventh adds kinematic redundancy for obstacle and joint-limit avoidance. Grübler's formula relates the DOF of a linkage to its joints and links. Why it matters for physical AI: DOF count determines what poses a robot can reach, how it handles singularities, and the dimensionality of its action space. Learned policies must scale with DOF, which is why humanoids and dexterous hands are markedly harder than 6-DOF arms. ### Delivery Robot Category: Applications URL: https://almond.bot/glossary/delivery-robot Delivery Robot is an autonomous mobile robot that transports goods over the last meters or miles of a logistics chain, spanning sidewalk robots, indoor hotel and hospital couriers, and delivery drones. These systems integrate localization, pedestrian-aware navigation, secure cargo compartments, and remote teleoperation fallback. Commercial fleets have operated on university campuses and city sidewalks since the late 2010s. Why it matters for physical AI: Delivery fleets are among the largest real-world deployments of autonomous navigation, generating operating data at scale and stress-testing perception, remote supervision, and regulatory frameworks that broader mobile robotics builds upon. ### Delta Robot Category: Hardware URL: https://almond.bot/glossary/delta-robot Delta Robot is a parallel-kinematic robot in which three arms with parallelogram linkages connect base-mounted actuators to a common end-effector platform, invented by Reymond Clavel at EPFL in the 1980s. Keeping motors on the stationary base minimizes moving mass, enabling very high accelerations. Delta robots dominate high-speed pick-and-place in food, pharmaceutical, and electronics packaging, and the architecture also appears in 3D printers. Why it matters for physical AI: Deltas illustrate how parallel kinematics trades workspace volume for speed and stiffness, a design point learned pick-and-place systems increasingly target. Their sub-second cycle times set the benchmark that vision-guided sorting must meet in production. ### Demonstration Category: Robot Learning URL: https://almond.bot/glossary/demonstration Demonstration is a recorded trajectory of observation-action pairs showing how a task should be performed, serving as the supervision signal for imitation learning. Demonstrations are gathered through teleoperation, kinesthetic teaching in which the arm is physically guided, or retargeted from human video. Quality dimensions include smoothness, consistency across operators, coverage of recovery behaviors, and alignment between the demonstrator's viewpoint and the robot's sensors. Why it matters for physical AI: Demonstration collection is the dominant cost in training manipulation policies, and demonstration quality frequently bounds final performance. Scalable, ergonomic collection interfaces are consequently a competitive frontier in physical AI. ### Denavit-Hartenberg Parameters Category: Math & Kinematics URL: https://almond.bot/glossary/denavit-hartenberg-parameters Denavit-Hartenberg Parameters are a minimal convention, introduced by Jacques Denavit and Richard Hartenberg in 1955, for describing the geometry of a serial kinematic chain. Each joint-link pair is characterized by four quantities, link length, link twist, link offset, and joint angle, defining the homogeneous transform between consecutive frames. Chaining these transforms yields the manipulator's forward kinematics; modified DH variants fix ambiguities in the original formulation. Why it matters for physical AI: DH parameters remain the lingua franca for specifying arm geometry in datasheets, calibration routines, and kinematics libraries. Accurate kinematic models built on them underpin the pose accuracy that learned controllers and calibration pipelines assume. ### Dense Optical Flow Category: Perception URL: https://almond.bot/glossary/dense-optical-flow Dense Optical Flow is the per-pixel estimation of apparent motion between consecutive images, producing a vector field over the entire frame rather than at sparse feature points. Classical variational methods such as Horn-Schunck have largely been superseded by learned estimators, with RAFT (2020) establishing the modern recurrent refinement paradigm. Flow supports motion segmentation, visual odometry, and video-based dynamics understanding. Why it matters for physical AI: Motion cues let robots segment moving objects, estimate ego-motion, and track manipulated items without task-specific labels. Flow-based representations also serve as intermediate action abstractions in some manipulation policies trained from video. ### Depth Anything Category: Perception URL: https://almond.bot/glossary/depth-anything Depth Anything is a family of monocular depth estimation foundation models, first released in 2024, trained on a combination of labeled datasets and tens of millions of unlabeled images via large-scale pseudo-labeling. It produces robust relative depth across diverse scenes, with fine-tuned variants providing metric depth. Depth Anything V2 improved detail and robustness using synthetic labels and larger teacher models. Why it matters for physical AI: Reliable depth from a single RGB camera reduces sensor cost and calibration burden on robot platforms. Foundation-model depth priors increasingly feed grasping, navigation, and 3D scene reconstruction pipelines without per-robot training. ### Depth Camera Category: Hardware URL: https://almond.bot/glossary/depth-camera Depth Camera is a sensor that measures per-pixel distance to the scene, producing depth maps typically alongside registered RGB images. Principal technologies include structured light, time-of-flight, and active or passive stereo, commercialized in devices such as the Microsoft Kinect and Intel RealSense series. Each modality trades range, resolution, and robustness differently; most struggle with transparent, reflective, or strongly sunlit surfaces. Why it matters for physical AI: Depth sensing gives manipulation and navigation systems direct 3D geometry for grasp planning, obstacle avoidance, and scene reconstruction. Sensor artifacts on glass and metal remain a practical failure source that perception stacks must handle in deployment. ### Depth Completion Category: Perception URL: https://almond.bot/glossary/depth-completion Depth Completion is the task of producing a dense depth map from sparse or incomplete depth measurements, usually guided by an aligned RGB image. Sparsity arises from LiDAR's limited scan lines, stereo failures on textureless regions, and depth-camera dropouts on reflective or transparent surfaces. Modern methods use convolutional or transformer networks that fuse image features with the valid depth samples to inpaint missing regions. Why it matters for physical AI: Grasp planners and collision checkers fail on holes in depth data, which are common precisely on the shiny and transparent objects robots must handle. Completion recovers usable geometry from imperfect, low-cost sensors. ### Depth Estimation Category: Perception URL: https://almond.bot/glossary/depth-estimation Depth Estimation is the inference of scene distance from camera imagery, either by triangulating correspondences across calibrated stereo pairs or by predicting depth from a single image with a learned model. Monocular methods, from early supervised networks to foundation models like MiDaS and Depth Anything, exploit pictorial cues and large-scale training to generalize across scenes, though they typically recover depth only up to an unknown scale unless metrically fine-tuned. Why it matters for physical AI: Cameras are the cheapest, densest sensors on a robot, and turning their pixels into 3D structure enables navigation and manipulation without dedicated depth hardware. Scale ambiguity and calibration remain the key deployment caveats. ### DexCap Category: Data & Benchmarks URL: https://almond.bot/glossary/dexcap DexCap is a portable hand motion capture system for collecting dexterous manipulation data directly from humans, introduced by Stanford researchers in 2024. It pairs electromagnetic finger-tracking gloves with chest-mounted cameras providing SLAM-based 6-DOF wrist tracking, recording precise hand poses alongside egocentric 3D observations in everyday environments. The paired DexIL framework retargets this data to train policies for multi-fingered robot hands without teleoperation. Why it matters for physical AI: Teleoperating high-DOF hands is slow and unnatural, bottlenecking dexterous datasets. Wearable capture systems let humans generate manipulation data at natural speed in the wild, a promising route to scaling beyond lab teleoperation. ### Dexterity Category: Manipulation URL: https://almond.bot/glossary/dexterity Dexterity is a robot's capacity to skillfully control interaction with objects, encompassing precise finger and arm coordination, regulation of contact forces, in-hand repositioning, and tool use. It is often characterized by the variety of grasp types and manipulation primitives a system can execute and the object diversity it tolerates. Human-level dexterity combines roughly two dozen hand DOF with dense tactile feedback, a combination robots have not yet matched. Why it matters for physical AI: Most economically valuable manual work, from assembly to food preparation, demands dexterity beyond parallel-jaw pick-and-place. Closing the dexterity gap is widely viewed as a defining milestone for general-purpose robots. ### Dexterous Hand Category: Hardware URL: https://almond.bot/glossary/dexterous-hand Dexterous Hand is a multi-fingered robotic end-effector with many actuated degrees of freedom, designed to approximate the grasping and in-hand manipulation versatility of the human hand. Notable examples include the Shadow Dexterous Hand with over 20 DOF, the Allegro Hand, and low-cost open-source designs such as the LEAP Hand. Key design tensions involve actuation (tendon-driven versus direct), tactile sensor integration, robustness, and cost. Why it matters for physical AI: Anthropomorphic hands unlock tasks and tools designed for humans, and their affordability increasingly determines whether dexterous manipulation research translates into fleets. Hardware fragility and sensing density remain the main obstacles to deployment. ### Dexterous Manipulation Category: Manipulation URL: https://almond.bot/glossary/dexterous-manipulation Dexterous Manipulation is the control of objects using coordinated multi-contact interaction, typically with multi-fingered hands, including in-hand reorientation, regrasping, and tool use. It requires reasoning over rapidly changing contact modes, friction, and rolling and sliding constraints. OpenAI's 2019 solution of a Rubik's Cube with the Shadow Hand, trained with reinforcement learning and automatic domain randomization, remains a landmark demonstration. Why it matters for physical AI: Reorienting a part before insertion or manipulating a tool mid-grasp separates genuinely useful manipulation from simple transport. Advances here directly expand the task repertoire of humanoids and dual-arm platforms in unstructured settings. ### Differential Drive Category: Locomotion URL: https://almond.bot/glossary/differential-drive Differential Drive is a mobile robot drive configuration using two independently controlled wheels on a common axis, with casters for balance. The robot moves straight when the wheels match speed, arcs when they differ, and turns in place when they counter-rotate. The platform is nonholonomic, unable to translate sideways, and its odometry follows directly from wheel encoder kinematics, making it the standard architecture for indoor mobile robots. Why it matters for physical AI: Most warehouse AMRs, service robots, and research bases are differential drive, so its kinematic constraints shape the motion planners and velocity controllers the whole mobile robotics stack assumes. ### Differential Flatness Category: Control URL: https://almond.bot/glossary/differential-flatness Differential Flatness is a system property in which all states and inputs can be expressed algebraically in terms of a set of flat outputs and finitely many of their time derivatives. For flat systems, trajectory planning reduces to designing sufficiently smooth curves in the flat output space, with the required control inputs recovered analytically. The quadrotor is the canonical robotic example, with position and yaw as flat outputs, a result exploited by Mellinger and Kumar's minimum-snap trajectory generation. Why it matters for physical AI: Flatness turns aggressive trajectory generation into tractable spline optimization, enabling the agile drone flight and smooth mobile-base motion that reactive learned planners are often benchmarked against. ### Differential Inverse Kinematics Category: Math & Kinematics URL: https://almond.bot/glossary/differential-inverse-kinematics Differential Inverse Kinematics is the velocity-level resolution of inverse kinematics, mapping desired end-effector twists to joint velocities through the manipulator Jacobian. Practical solvers use the pseudoinverse or damped least squares to remain stable near singularities, and quadratic-programming formulations add joint limits, collision constraints, and secondary objectives such as posture optimization for redundant arms. It underlies resolved-rate control and most real-time Cartesian servoing. Why it matters for physical AI: Learned policies frequently output end-effector velocities or pose targets, delegating joint-space execution to differential IK. Its conditioning near singularities and constraint handling determine whether those commands translate into safe, smooth hardware motion. ### Diffusion Model Category: Foundation Models URL: https://almond.bot/glossary/diffusion-model Diffusion Model is a generative model that learns to reverse a gradual noising process, generating samples by iteratively denoising from pure noise. Formalized in denoising diffusion probabilistic models (Ho et al., 2020) and score-based frameworks, diffusion models achieve state-of-the-art image, video, and trajectory synthesis with stable training and strong mode coverage. Sampling acceleration techniques such as DDIM and distillation reduce the number of denoising steps required. Why it matters for physical AI: Beyond powering diffusion policies for control, these models generate synthetic training scenes, forecast video futures for world models, and synthesize trajectories, making them a core generative engine across the physical AI stack. ### Diffusion Policy Category: Robot Learning URL: https://almond.bot/glossary/diffusion-policy Diffusion Policy is a visuomotor policy class that generates robot action sequences by iteratively denoising samples conditioned on visual observations, introduced by Chi et al. in 2023. Modeling the action distribution generatively lets it capture multimodal demonstrations that regression-based behavior cloning averages into invalid actions, and predicting action chunks over a receding horizon yields temporally consistent motion. It substantially outperformed prior imitation methods across simulated and real manipulation benchmarks. Why it matters for physical AI: Diffusion-based action heads have become a default choice for demonstration-trained manipulation and appear inside several robot foundation models. Their inference latency motivates ongoing work on distillation and faster samplers for high-rate control. ### Digit Category: Hardware URL: https://almond.bot/glossary/digit Digit is a bipedal humanoid robot developed by Agility Robotics for logistics work, evolved from the company's Cassie biped, which originated in legged locomotion research at Oregon State University. Digit combines dynamically stable walking on bird-inspired legs with an upper torso and arms for tote handling, and has been piloted in warehouse deployments moving containers between conveyors and shelving. Why it matters for physical AI: Digit is among the first humanoids operated in commercial pilots rather than demos, making it a reference point for the economics, safety cases, and reliability engineering of humanoid labor. ### Digital Manufacturing Category: Applications URL: https://almond.bot/glossary/digital-manufacturing Digital Manufacturing is the integrated use of computer-based systems, including simulation, 3D modeling, analytics, and connected machines, to design, plan, and execute production. It links product definition to factory-floor equipment such as CNC machines and industrial robots, enabling virtual commissioning, process optimization, and rapid changeover. The approach underpins Industry 4.0 initiatives and increasingly incorporates digital twins of entire production lines. Why it matters for physical AI: Factories instrumented for digital manufacturing provide the structured data, simulation assets, and connectivity that learning-based robots need for fast integration, making them the most tractable early market for adaptive automation. ### Digital Twin Category: Simulation URL: https://almond.bot/glossary/digital-twin Digital Twin is a virtual replica of a physical asset, process, or environment that is kept synchronized with its real counterpart through sensor data and state updates. In robotics, digital twins span calibrated simulation models of individual robots to full work cells and warehouses, supporting virtual commissioning, what-if analysis, monitoring, and policy training. Fidelity requirements differ by use: control-grade twins demand accurate dynamics, while planning twins prioritize geometry and throughput statistics. Why it matters for physical AI: High-fidelity twins let operators validate robot behavior and train or evaluate policies against a faithful stand-in before touching production, shrinking the sim-to-real gap and de-risking deployment changes. ### DINOv2 Category: Foundation Models URL: https://almond.bot/glossary/dinov2 DINOv2 is a family of self-supervised Vision Transformers released by Meta AI in 2023, trained with a discriminative self-distillation objective on a curated dataset of 142 million images. Its frozen features transfer strongly to classification, segmentation, depth estimation, and dense correspondence without fine-tuning. The 2025 successor DINOv3 scaled the approach further. Why it matters for physical AI: Frozen self-supervised backbones like DINOv2 are widely used as visual encoders for manipulation policies and correspondence-based grasping, providing robust, semantically rich features from limited robot data. ### Direct Drive Actuator Category: Hardware URL: https://almond.bot/glossary/direct-drive-actuator Direct Drive Actuator is an actuator that couples the motor to the load without a gearbox, or in quasi-direct-drive form through a low single-digit reduction. Eliminating high gear ratios removes backlash and reflected inertia, yielding high backdrivability and accurate torque estimation from motor current, at the cost of requiring large-diameter, high-torque motors. The MIT Mini Cheetah popularized quasi-direct-drive actuation for dynamic legged robots. Why it matters for physical AI: Transparent, backdrivable actuation enables proprioceptive contact detection and safe physical interaction without dedicated force sensors, properties that dynamic locomotion and contact-rich manipulation policies increasingly rely on. ### Direct Teaching Category: Teleoperation URL: https://almond.bot/glossary/direct-teaching Direct Teaching is the programming of a robot by physically guiding it through desired motions, also called kinesthetic teaching or hand-guiding, with the trajectory recorded for replay or as training data. It requires gravity-compensated, backdrivable, or force-sensing hardware so the arm follows the human's hand, and is a standard commissioning mode on collaborative robots. ISO/TS 15066 covers hand-guided operation as a collaborative scenario. Why it matters for physical AI: Hand-guiding lets non-programmers produce demonstrations in minutes, lowering the skill barrier for both classical waypoint programming and imitation learning data collection on factory floors. ### Discriminator Category: Robot Learning URL: https://almond.bot/glossary/discriminator Discriminator is the component of an adversarial learning system trained to distinguish real data from generated or agent-produced data, as introduced in generative adversarial networks (Goodfellow et al., 2014). In robot learning, discriminators serve as learned reward functions: Generative Adversarial Imitation Learning (GAIL) rewards the agent for producing state-action pairs the discriminator cannot tell apart from expert demonstrations, and Adversarial Motion Priors apply the same idea to naturalistic locomotion styles. Why it matters for physical AI: Adversarially learned rewards sidestep manual reward engineering, letting legged robots and characters acquire lifelike, demonstration-consistent behavior from modest motion datasets. ### Disparity Map Category: Perception URL: https://almond.bot/glossary/disparity-map Disparity Map is an image encoding, for each pixel, the horizontal offset between its projections in a rectified stereo pair. Disparity is inversely proportional to depth given the cameras' focal length and baseline, so an accurate disparity map is equivalent to a metric depth map. Classical block matching and semi-global matching have been largely superseded by learned stereo networks, though textureless and reflective regions remain difficult. Why it matters for physical AI: Stereo disparity provides metrically scaled depth from passive cameras, a robust and inexpensive geometry source for outdoor navigation and manipulation where active depth sensors wash out or interfere. ### Distributional Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/distributional-reinforcement-learning Distributional Reinforcement Learning is a family of value-based methods that model the full probability distribution of returns rather than only their expectation. Introduced with the C51 algorithm by Bellemare, Dabney, and Munos in 2017 and extended by quantile regression approaches such as QR-DQN and IQN, distributional critics often improve learning stability and enable risk-sensitive action selection based on return quantiles. Why it matters for physical AI: Risk-aware control matters when failures break hardware or endanger people. Return distributions let deployed policies act conservatively in high-variance situations instead of optimizing average-case outcomes. ### Dobb-E Category: Robot Learning URL: https://almond.bot/glossary/dobb-e Dobb-E is an open-source framework from NYU (2023) for teaching mobile manipulators household tasks from brief human demonstrations. Data is collected with the Stick, a reacher-grabber tool carrying an iPhone, yielding the Homes of New York dataset of thousands of demonstrations across dozens of real homes. A visual representation pretrained on this data lets a Hello Robot Stretch learn new home tasks from roughly five minutes of demonstrations with high success rates. Why it matters for physical AI: Cheap, phone-based data collection in real homes demonstrated a practical alternative to lab teleoperation for scaling household manipulation data, influencing subsequent low-cost data-gathering tool designs. ### Docker in Robotics Category: Software & Middleware URL: https://almond.bot/glossary/docker-in-robotics Docker in Robotics is the use of container technology to package robot software with its exact dependencies into portable, reproducible images. Containers isolate conflicting ROS distributions, CUDA versions, and Python environments, standardize CI testing, and enable fleet-wide deployment and rollback of robot software. On hardware, containers typically run with device passthrough for GPUs, cameras, and serial buses, and with host networking for DDS discovery. Why it matters for physical AI: Reproducible environments are what allow a policy validated in CI to behave identically on hundreds of robots. Containerized deployment underpins over-the-air updates and scalable fleet operations. ### Domain Adaptation Category: Robot Learning URL: https://almond.bot/glossary/domain-adaptation Domain Adaptation is a set of transfer learning techniques for making a model trained on a source data distribution perform well on a different target distribution, typically with little or no labeled target data. Approaches include adversarial feature alignment, image-to-image translation such as CycleGAN-based sim-to-real pixel adaptation, and self-training on pseudo-labels. In robotics the canonical source-target pairs are simulation versus reality and one deployment site versus another. Why it matters for physical AI: Every new cell, warehouse, or home shifts lighting, textures, and object distributions away from the training set. Adaptation techniques reduce the recollection and retraining cost each shift would otherwise incur. ### Domain Gap Category: Simulation URL: https://almond.bot/glossary/domain-gap Domain Gap is the mismatch between the data distribution a model was trained on and the distribution it encounters at deployment, causing performance degradation. In robotics the most studied instance is the sim-to-real gap, spanning visual discrepancies in rendering, lighting, and textures as well as physical discrepancies in contact dynamics, actuation latency, and sensor noise. Gaps also arise between labs and deployment sites, and between embodiments. Why it matters for physical AI: The size of the domain gap determines how much simulation training, offline data, and cross-robot data actually transfer to a target platform, making gap measurement and mitigation central to scaling physical AI economically. ### Domain Randomization Category: Simulation URL: https://almond.bot/glossary/domain-randomization Domain Randomization is a sim-to-real technique that randomizes simulator parameters, such as textures, lighting, camera poses, object properties, and physics coefficients, during training so the learned policy or perception model treats reality as one more variation. Popularized by Tobin et al. in 2017 for vision transfer, and extended by automatic domain randomization in OpenAI's dexterous manipulation work, it trades peak in-domain performance for robustness across the randomized range. Why it matters for physical AI: Randomization remains the most reliable recipe for transferring simulation-trained locomotion and manipulation policies to hardware without real-world fine-tuning, and calibrating its ranges is a standing engineering craft in sim-to-real pipelines. ### Drake Category: Software & Middleware URL: https://almond.bot/glossary/drake Drake is an open-source C++ and Python toolbox for model-based design and verification of robotic systems, initiated by Russ Tedrake's group at MIT and developed with the Toyota Research Institute. It provides careful multibody dynamics with rigorous contact modeling, a systems framework for composing controllers and estimators, and bindings to mathematical programming solvers for trajectory optimization and control synthesis. Drake emphasizes numerical accuracy and gradients over raw simulation speed. Why it matters for physical AI: Optimization-based planning and contact-implicit trajectory synthesis complement learned policies, and Drake is a standard research platform for that model-based half of modern manipulation and locomotion stacks. ### DrEureka Category: Robot Learning URL: https://almond.bot/glossary/dreureka DrEureka is a 2024 method that uses large language models to automate sim-to-real transfer design, extending the Eureka reward-generation framework. Given a task and simulator source code, the LLM proposes reward functions and, guided by a physics-feasibility prior, generates domain randomization ranges for training robust policies. Its headline demonstration was a quadruped balancing and walking on a yoga ball, transferred to hardware zero-shot without task-specific manual tuning. Why it matters for physical AI: Reward shaping and randomization tuning are the most labor-intensive parts of sim-to-real reinforcement learning. LLM-automated pipeline design points toward compressing weeks of expert iteration into hours of compute. ### DROID Category: Data & Benchmarks URL: https://almond.bot/glossary/droid DROID (Distributed Robot Interaction Dataset) is a large-scale manipulation dataset released in 2024, containing roughly 76,000 teleoperated demonstration trajectories, about 350 hours, collected on a standardized Franka arm platform by dozens of institutions across North America, Europe, and Asia. Each episode includes multi-view stereo camera streams, language annotations, and full controller states, with deliberate diversity in scenes, tasks, and objects across hundreds of buildings. Why it matters for physical AI: Policies co-trained on DROID show measurably better robustness to novel scenes and objects, evidence that standardized, in-the-wild data collection improves generalization in ways lab-only datasets cannot. ### Drone Category: Applications URL: https://almond.bot/glossary/drone Drone is the common term for an unmanned aerial vehicle (UAV), an aircraft operated without an onboard pilot, ranging from palm-sized quadrotors to fixed-wing survey aircraft. Multirotor drones combine IMU-based attitude stabilization, GPS or visual-inertial state estimation, and electric propulsion, and serve applications from aerial imaging and inspection to mapping, agriculture, and delivery. Research platforms have demonstrated agile autonomous flight, including drone racing at champion human level with the Swift system. Why it matters for physical AI: Drones are the most mature mass-market autonomous robots, and their perception, state estimation, and swarm coordination advances feed directly into broader physical AI, while aerial data collection supports mapping and inspection workflows. ### Drone Swarm Category: Applications URL: https://almond.bot/glossary/drone-swarm Drone Swarm is a group of unmanned aerial vehicles coordinating to accomplish a shared objective, using either centralized planning or decentralized control in which each vehicle acts on local sensing and neighbor communication. Research milestones include thousand-drone light shows with precomputed trajectories and decentralized swarms flying through forests, such as the Zhejiang University micro-swarm work published in 2022. Core problems include inter-agent collision avoidance, relative localization, and communication-constrained planning. Why it matters for physical AI: Swarms trade individual capability for redundancy and coverage, a pattern relevant to warehouse fleets and inspection teams. Their decentralized coordination algorithms transfer to any multi-robot deployment operating under limited bandwidth. ### Dropout Category: Robot Learning URL: https://almond.bot/glossary/dropout Dropout is a neural network regularization technique that randomly zeroes a fraction of unit activations during each training step, preventing co-adaptation of features and reducing overfitting, introduced by Srivastava, Hinton, and colleagues in 2014. At inference, all units are active with appropriately scaled weights. Monte Carlo dropout, which keeps sampling at test time, provides an inexpensive approximation of Bayesian predictive uncertainty. Why it matters for physical AI: Robot datasets are small relative to model capacity, making regularization consequential for policy generalization. Monte Carlo dropout additionally offers a cheap uncertainty signal usable for failure detection and human handover triggers. ### Dual Numbers Category: Math & Kinematics URL: https://almond.bot/glossary/dual-numbers Dual Numbers are an extension of the real numbers of the form a + bε, where ε is a nilpotent element satisfying ε² = 0. Evaluating a function on a dual number yields its derivative in the dual component, which is the mechanism behind forward-mode automatic differentiation. In kinematics, dual numbers combine with quaternions to form dual quaternions, and with screw theory they compactly encode coupled rotation and translation. Why it matters for physical AI: Automatic differentiation built on dual-number arithmetic supplies the exact gradients used in differentiable simulation, calibration, and trajectory optimization, while dual-quaternion kinematics gives singularity-free pose representations for control. ### Dual Quaternion Category: Math & Kinematics URL: https://almond.bot/glossary/dual-quaternion Dual Quaternion is an eight-parameter algebraic representation of rigid-body transformations, formed as a quaternion with dual-number coefficients, encoding rotation and translation in a single object. Unit dual quaternions compose transformations by multiplication, interpolate smoothly along screw motions, and avoid the singularities of Euler angles. They are used in kinematic modeling, hand-eye calibration, pose-graph optimization, and skinning in graphics. Why it matters for physical AI: Compact, singularity-free pose algebra matters wherever transforms are estimated or interpolated, from calibration routines to pose regression targets in learned perception, where representation choice measurably affects accuracy. ### Dual-Arm Robot Category: Hardware URL: https://almond.bot/glossary/dual-arm-robot Dual-Arm Robot is a system with two coordinated manipulators, enabling bimanual skills such as stabilizing an object with one arm while the other acts on it, handling large or deformable items, and performing human-style assembly. Control challenges include closed kinematic chains when both arms grasp one object, inter-arm collision avoidance, and role allocation. Bimanual teleoperation systems such as ALOHA have made dual-arm demonstration collection dramatically cheaper. Why it matters for physical AI: Many everyday tasks, folding laundry, opening containers, connecting cables, are inherently bimanual, and dual-arm platforms paired with imitation learning have become a leading testbed for generalist manipulation policies. ### Duty Cycle Category: Hardware URL: https://almond.bot/glossary/duty-cycle Duty Cycle is the fraction of a period during which a signal or system is active. In pulse-width modulation, the duty cycle of a fixed-frequency square wave sets the average voltage delivered to a motor, and thus its torque or speed. The term also denotes an actuator's rated on-time fraction, the proportion of continuous operation a motor or robot can sustain without exceeding thermal limits. Why it matters for physical AI: PWM duty cycles are the lowest-level interface between control software and actuators, while thermal duty ratings bound how hard learned controllers can drive hardware in sustained deployment without derating or failure. ### Dynamic Gait Category: Locomotion URL: https://almond.bot/glossary/dynamic-gait Dynamic Gait is a legged locomotion pattern that includes phases where the robot is not statically stable, relying on momentum and continuous active control to avoid falling, in contrast to static gaits that keep the center of mass over the support polygon at all times. Trotting, bounding, galloping, and human-like walking with aerial or two-legged support phases are dynamic. Analysis tools include the zero-moment point and capture point; modern quadrupeds execute dynamic gaits via model predictive control or reinforcement-learned policies. Why it matters for physical AI: Speed and efficiency on real terrain require dynamic gaits, and their reliability on stairs, gravel, and slopes has become the benchmark by which learned locomotion controllers are judged for deployment. ### Dynamic Grasping Category: Manipulation URL: https://almond.bot/glossary/dynamic-grasping Dynamic Grasping is the grasping of objects while the object, the robot, or both are in motion, such as picking items from a moving conveyor, catching thrown objects, or grasping from a moving mobile base. It couples real-time object pose prediction with reactive trajectory generation so the gripper intercepts the object within its closing window. Demonstrations include high-speed robot catching of balls and bottles using motion-capture-based prediction. Why it matters for physical AI: Conveyor picking and mobile manipulation both demand grasping under motion, and the tight perception-to-action latency budgets involved stress-test the reactive capabilities of learned visuomotor policies. ### Dynamic Movement Primitives (DMP) Category: Control URL: https://almond.bot/glossary/dynamic-movement-primitives Dynamic Movement Primitives (DMP) are a framework for encoding motions as stable nonlinear dynamical systems, developed by Ijspeert, Nakanishi, and Schaal in the early 2000s. A spring-damper attractor guarantees convergence to a goal while a learned forcing term, fit from a single demonstration, shapes the trajectory; the formulation supports temporal scaling, goal changes, and rhythmic as well as discrete motions. DMPs were a workhorse of pre-deep-learning imitation learning. Why it matters for physical AI: DMPs remain attractive when guarantees matter: one demonstration suffices, goal convergence is provable, and trajectories adapt online, properties that black-box policy networks are still working to match in safety-conscious settings. ### Dynamic Window Approach (DWA) Category: Navigation & SLAM URL: https://almond.bot/glossary/dynamic-window-approach Dynamic Window Approach (DWA) is a local obstacle-avoidance and velocity-planning method for mobile robots, introduced by Fox, Burgard, and Thrun in 1997. It samples candidate translational and rotational velocities from the window reachable within one control cycle given acceleration limits, simulates short-horizon arcs, discards colliding ones, and scores the rest on heading, clearance, and speed. DWA has long been a standard local planner in ROS navigation stacks. Why it matters for physical AI: Decades after publication, DWA-style local planners still execute the final meters of most AMR motion, forming the reactive safety layer beneath global planners and, increasingly, learned navigation policies. ### Dynamics Model Category: Control URL: https://almond.bot/glossary/dynamics-model Dynamics Model is a mathematical or learned representation of how a system's state evolves under applied inputs. Analytic robot dynamics take the manipulator equation form derived from Lagrangian mechanics, with mass matrix, Coriolis, and gravity terms, while learned dynamics models are neural networks trained on transition data, as in model-based reinforcement learning and world models. Accuracy over long prediction horizons, especially through contact events, is the central difficulty. Why it matters for physical AI: A usable dynamics model turns control into planning: model predictive control, trajectory optimization, and model-based RL all trade real-world trial and error for computation against the model, a decisive advantage when hardware time is scarce. ### Dynamics Randomization Category: Simulation URL: https://almond.bot/glossary/dynamics-randomization Dynamics Randomization is the randomization of physical simulation parameters, including link masses, friction coefficients, motor strength, damping, latency, and sensor noise, during policy training so the learned controller is robust to the true system's unknown dynamics. Peng et al. (2018) showed it enables sim-to-real transfer of manipulation policies without real-world fine-tuning, and it is standard in learned quadruped and humanoid locomotion pipelines, sometimes paired with online system identification. Why it matters for physical AI: Contact dynamics and actuator behavior are the hardest quantities to simulate faithfully, so randomizing over them is often the difference between a policy that walks on hardware and one that falls immediately. ### Dynamixel Category: Hardware URL: https://almond.bot/glossary/dynamixel Dynamixel is a line of integrated smart servo actuators made by the Korean company Robotis, combining a DC motor, gearbox, encoder, controller, and daisy-chainable serial bus interface in one package. Servos report position, velocity, load, and temperature, and support position, velocity, and current control modes. Their low cost and simple integration have made them ubiquitous in research arms, hobby humanoids, and low-cost teleoperation devices, including leader arms in ALOHA-style bimanual rigs. Why it matters for physical AI: Affordable integrated servos drove the sub-thousand-dollar arm ecosystem that democratized imitation learning research, putting demonstration collection hardware within reach of individual labs and hobbyists. ## Glossary: E ### Edge Computing Category: Software & Middleware URL: https://almond.bot/glossary/edge-computing Edge Computing is the execution of computation on or near the device generating the data rather than in remote cloud datacenters. For robots this means running perception, planning, and policy inference on onboard embedded GPUs and accelerators, such as NVIDIA's Jetson line, or on nearby servers. It bounds control latency, preserves operation through network outages, and keeps sensitive sensor data local. Why it matters for physical AI: Closed-loop control cannot tolerate cloud round-trip latency or dropped connectivity, so model efficiency on embedded hardware directly constrains which foundation models can actually run on deployed robots. ### Edge Detection Category: Perception URL: https://almond.bot/glossary/edge-detection Edge Detection is the identification of pixels where image intensity changes sharply, marking object boundaries, surface creases, and texture transitions. Classical operators include the Sobel and Laplacian filters and the Canny detector (1986), which combines Gaussian smoothing, gradient computation, non-maximum suppression, and hysteresis thresholding. Learned detectors and the edge-sensitive features inside modern convolutional networks have absorbed much of its standalone role. Why it matters for physical AI: Edges remain a lighting-robust cue for part localization, weld seam tracking, and fixture alignment in industrial vision, and edge-like features are among the first representations visual encoders learn. ### Ego4D Category: Data & Benchmarks URL: https://almond.bot/glossary/ego4d Ego4D is a large-scale egocentric video dataset and benchmark suite released in 2022 by Meta AI and a consortium of universities, containing roughly 3,670 hours of daily-life first-person video from over 900 camera wearers across many countries. It includes annotations for episodic memory, hand-object interaction, forecasting, and social understanding tasks. Ego4D has become a standard pretraining corpus for egocentric perception and video representation models. Why it matters for physical AI: First-person human video is the closest abundant proxy for robot manipulation viewpoints, and representations pretrained on Ego4D, such as R3M, transfer measurably to robot policy learning. ### Egocentric Video Category: Data & Benchmarks URL: https://almond.bot/glossary/egocentric-video Egocentric Video is first-person footage captured from a camera worn on the head, chest, or glasses of the person performing an activity, in contrast to third-person views of the actor. It exhibits characteristic hand-centric framing, motion blur from head movement, and tight coupling between gaze and action. Datasets such as Ego4D and Epic-Kitchens established benchmarks for hand-object interaction understanding and action anticipation from this viewpoint. Why it matters for physical AI: Wearable cameras can capture human manipulation at a scale no robot fleet matches, and the egocentric viewpoint roughly aligns with wrist- and head-mounted robot cameras, making such video a key ingredient for pretraining manipulation policies. ### EgoMimic Category: Robot Learning URL: https://almond.bot/glossary/egomimic EgoMimic is a 2024 framework from Georgia Tech for scaling manipulation learning via egocentric human video, using Meta's Project Aria smart glasses to record humans performing tasks and a bimanual robot with a matched camera viewpoint. It co-trains policies on human embodiment data and robot demonstrations, bridging the gap with hand-to-gripper action retargeting and normalized visual observations, improving performance and generalization over robot-data-only baselines. Why it matters for physical AI: Treating human video as first-class training data rather than pretraining material points to a path where an hour of wearing glasses substitutes for an hour of expensive robot teleoperation. ### Elastic Actuation Category: Hardware URL: https://almond.bot/glossary/elastic-actuation Elastic Actuation is the deliberate introduction of compliance, typically a spring, between an actuator's gearbox and its load. The canonical form is the series elastic actuator, introduced by Pratt and Williamson in 1995, in which measuring spring deflection turns force control into a position-sensing problem, providing clean force regulation, shock tolerance, and energy storage. Variable stiffness actuators extend the idea with mechanically adjustable compliance. Why it matters for physical AI: Compliance protects gearboxes from impact and makes physical human-robot contact intrinsically safer, which is why elastic actuation appears in exoskeletons, legged robots, and collaborative platforms designed to work around people. ### Eldercare Robot Category: Applications URL: https://almond.bot/glossary/eldercare-robot Eldercare Robot is a robot designed to support older adults, spanning social companions such as PARO, telepresence and monitoring platforms, mobility and transfer assistance devices, and mobile manipulators for fetching and household tasks. Deployment demands reliability around fragile users, intuitive interfaces for non-technical operators, and privacy-conscious sensing. Aging demographics, most acutely in Japan, have made eldercare a long-standing driver of assistive robotics research. Why it matters for physical AI: Demographic pressure makes eldercare one of the clearest large markets for general-purpose home robots, while its safety, dignity, and trust requirements set the hardest bar for deploying learned manipulation around people. ### Electronics Manufacturing Robot Category: Applications URL: https://almond.bot/glossary/electronics-manufacturing-robot Electronics Manufacturing Robot is a robot used in the production of electronic devices, including SMT pick-and-place machines populating circuit boards at tens of thousands of components per hour, SCARA and delta robots for assembly and testing, and precision cells for connector insertion, screwdriving, and adhesive dispensing. The domain demands micron-scale repeatability, electrostatic discharge compliance, and handling of small, delicate, often flexible parts such as ribbon cables. Why it matters for physical AI: Electronics assembly concentrates exactly the fine, contact-rich, high-mix tasks that motivate learned dexterity, and its enormous manual labor base makes it a prime target market for adaptive manipulation. ### Embedding Category: Foundation Models URL: https://almond.bot/glossary/embedding Embedding is a learned dense vector representation that places discrete or high-dimensional inputs, words, images, states, or skills, into a continuous space where geometric proximity reflects semantic similarity. Embeddings are produced by encoder networks and underpin retrieval, clustering, and multimodal alignment, as in CLIP's joint image-text space. In robotics, observation and language embeddings serve as the interface between perception, instructions, and policy networks. Why it matters for physical AI: Shared embedding spaces are what let a language instruction retrieve the right visual features and condition a manipulation policy, forming the connective tissue of vision-language-action systems. ### Embodied AI Category: Foundation Models URL: https://almond.bot/glossary/embodied-ai Embodied AI is the study of intelligent agents that perceive and act within a physical or physically simulated environment through a body, in contrast to disembodied systems that process static datasets. The field spans robotics and simulated benchmarks, navigation, rearrangement, and embodied question answering in platforms such as AI Habitat, and is grounded in the thesis that intelligence develops through sensorimotor interaction with the world. Why it matters for physical AI: The embodied framing drives the research agenda behind robot foundation models: internet-scale knowledge must be grounded in perception and action loops before it produces competent physical behavior. ### Embodiment Category: Foundation Models URL: https://almond.bot/glossary/embodiment Embodiment is the physical form through which an agent senses and acts, encompassing its morphology, kinematics, actuation, and sensor suite. An embodiment defines the agent's observation and action spaces, so the same task demands different policies on a quadruped, a wheeled base, or a bimanual humanoid. Cross-embodiment research, exemplified by the Open X-Embodiment collaboration, trains single models across many robot bodies. Why it matters for physical AI: Whether skills learned on one body transfer to another determines if robotics data can pool into shared foundation models or must be recollected per platform, one of the field's pivotal open questions. ### Embodiment Gap Category: Robot Learning URL: https://almond.bot/glossary/embodiment-gap Embodiment Gap is the mismatch in morphology, kinematics, sensing, and dynamics between the agent that generated training data and the robot executing the learned behavior. It arises most sharply when learning from human video, where hands, arm geometry, and viewpoints differ from grippers and robot cameras, and also between distinct robot platforms. Mitigations include action and viewpoint retargeting, embodiment-agnostic representations, and co-training on data from both embodiments. Why it matters for physical AI: Human video and heterogeneous robot fleets are the largest available data sources for manipulation, and the embodiment gap is precisely what limits how much of that data transfers to any given platform. ### Emergency Stop (E-Stop) Category: Safety & Standards URL: https://almond.bot/glossary/emergency-stop Emergency Stop (E-Stop) is a fail-safe function, typically triggered by a prominent red mushroom button, that immediately removes hazardous energy or halts machine motion, overriding all other controls. Its design is governed by ISO 13850 and IEC 60204-1, which require direct-opening contacts, latching until manual reset, and stop categories defining whether power is cut instantly or after a controlled deceleration. E-stop circuits are implemented in safety-rated hardware independent of application software. Why it matters for physical AI: Learned policies are treated as unverified software, so hardware safety layers like the e-stop remain the certified last line of defense that makes experimenting with and deploying such policies acceptable. ### Encoder Category: Hardware URL: https://almond.bot/glossary/encoder Encoder is a sensor that measures the angular or linear position of a mechanism, most commonly a rotary device on a motor shaft or joint using optical, magnetic, or capacitive sensing. Incremental encoders emit quadrature pulses counted relative to a reference, while absolute encoders report position uniquely on power-up. Joint encoders provide the proprioceptive feedback for servo control, and the term also names the input-compressing half of encoder-decoder neural architectures. Why it matters for physical AI: Every joint-level control loop and every proprioceptive observation fed to a learned policy originates in encoder readings, making their resolution and reliability foundational to both classical control and policy learning. ### Encoder Resolution Category: Hardware URL: https://almond.bot/glossary/encoder-resolution Encoder Resolution is the smallest position increment an encoder can distinguish, specified in counts or pulses per revolution for rotary devices, with quadrature decoding quadrupling the base line count. Resolution at the motor is multiplied through the gear ratio at the joint but so is the effect of backlash, and interpolated optical or magnetic encoders reach millions of counts per revolution. Resolution bounds velocity-estimation quality at low speeds. Why it matters for physical AI: Fine manipulation and smooth force estimation from motor current both hinge on high-resolution joint sensing, and resolution limits show up directly as noise in the proprioceptive streams robot learning datasets record. ### Encoder-Decoder Architecture Category: Foundation Models URL: https://almond.bot/glossary/encoder-decoder-architecture Encoder-Decoder Architecture is a neural network pattern in which an encoder compresses the input into a latent representation and a decoder expands that representation into the output. It spans sequence-to-sequence models with attention, the original transformer of Vaswani et al. (2017), U-Net's skip-connected variant for dense prediction, and autoencoders. The design separates understanding the input from generating the output, letting the two sides use different modalities or resolutions. Why it matters for physical AI: Robot policies are frequently encoder-decoder systems, vision and language encoders feeding action decoders, and diffusion policies use U-Net or transformer decoders, so the pattern's trade-offs shape policy architecture design. ### End-Effector Category: Hardware URL: https://almond.bot/glossary/end-effector End-Effector is the device mounted at the distal end of a robot arm that interacts with the environment, including parallel-jaw and multi-fingered grippers, suction cups, welding torches, screwdrivers, and cameras. It attaches at the tool flange, and its tool center point defines the frame in which Cartesian motions and task poses are specified. End-effector choice largely determines which tasks an arm can perform. Why it matters for physical AI: Gripper morphology constrains the entire learning problem: parallel jaws simplify data collection and policy output spaces, while richer end-effectors expand task coverage at the cost of harder control and data acquisition. ### End-Effector Pose Category: Math & Kinematics URL: https://almond.bot/glossary/end-effector-pose End-Effector Pose is the position and orientation of a manipulator's tool frame relative to a reference frame, an element of the group SE(3) with six degrees of freedom. It is computed from joint angles via forward kinematics and realized from a target via inverse kinematics, with orientation represented by rotation matrices, quaternions, or axis-angle forms. Pose accuracy depends on kinematic calibration and joint sensing. Why it matters for physical AI: Cartesian end-effector poses and velocities are the most common action space for learned manipulation policies, chosen because they transfer across arms with different kinematics far better than joint-space actions. ### End-of-Arm Tooling (EOAT) Category: Hardware URL: https://almond.bot/glossary/end-of-arm-tooling End-of-Arm Tooling (EOAT) is the industry term for the complete tooling assembly mounted on a robot's wrist flange, encompassing grippers, suction arrays, tool changers, force-torque sensors, cameras, and application-specific fixtures. EOAT is often custom-engineered per application and can rival the robot arm in cost. Standardized mechanical and electrical interfaces, such as ISO 9409 flange patterns and automatic tool changers, allow one arm to swap between tools mid-cycle. Why it matters for physical AI: Purpose-built tooling is how fixed automation achieves reliability, and reducing EOAT customization through general-purpose grippers plus adaptive policies is a core economic promise of learned manipulation. ### End-to-End Learning Category: Robot Learning URL: https://almond.bot/glossary/end-to-end-learning End-to-End Learning is the training of a single differentiable model that maps raw sensory input directly to control output, replacing hand-engineered pipelines of separate perception, state estimation, planning, and control modules. Levine et al.'s 2016 work on end-to-end visuomotor policies demonstrated pixels-to-torques manipulation, and the paradigm now dominates imitation-trained manipulation and vision-language-action models. Critiques center on interpretability, verification difficulty, and data hunger. Why it matters for physical AI: Whether robot autonomy should be one large trained network or a structured stack with learned components is the field's central architectural debate, with end-to-end approaches currently propelled by foundation model scaling results. ### Energy Shaping Category: Control URL: https://almond.bot/glossary/energy-shaping Energy Shaping is a nonlinear control methodology that modifies a system's apparent energy function so the desired behavior becomes a natural minimum of the closed-loop dynamics, with damping injection added to dissipate energy toward it. Formalized in passivity-based control and interconnection-and-damping-assignment (IDA-PBC), it exploits rather than cancels system dynamics. Classic robotic applications include swing-up of underactuated pendulums and compliant manipulation. Why it matters for physical AI: Controllers with energy-based stability certificates offer the formal guarantees that pure policy networks lack, and they increasingly serve as safe low-level layers underneath learned high-level commands on underactuated hardware. ### Energy-Based Model (EBM) Category: Robot Learning URL: https://almond.bot/glossary/energy-based-model Energy-Based Model (EBM) is a model that defines an unnormalized probability distribution through a scalar energy function, assigning low energy to compatible input configurations and high energy elsewhere. Inference is performed by minimizing energy or sampling, avoiding the need for a normalized likelihood. Implicit Behavioral Cloning (Florence et al., 2021) framed robot policies as EBMs over observation-action pairs, capturing multimodal and discontinuous action distributions that regression smooths over. Why it matters for physical AI: Multimodal demonstrations, where several actions are equally valid, break mean-regression policies. Energy-based and diffusion policies were the two responses that reshaped imitation learning architectures for manipulation. ### Epipolar Geometry Category: Perception URL: https://almond.bot/glossary/epipolar-geometry Epipolar Geometry is the intrinsic projective geometry relating two views of the same scene, encoded by the essential matrix for calibrated cameras and the fundamental matrix otherwise. A point in one image constrains its correspondence in the other to a line, the epipolar line, reducing stereo matching from a 2D to a 1D search. The essential matrix decomposes into the relative rotation and translation between cameras, the basis of two-view structure from motion. Why it matters for physical AI: Stereo depth, visual odometry, and multi-camera calibration on robots all rest on epipolar constraints, and they remain the geometric consistency check applied to learned correspondence and pose estimators. ### Episode Category: Robot Learning URL: https://almond.bot/glossary/episode Episode is a single complete interaction sequence between an agent and its environment, running from an initial state to termination by task success, failure, or a time limit, followed by a reset. In reinforcement learning, returns and success rates are computed per episode; in robot data collection, an episode typically corresponds to one demonstration of a task. Episode boundaries define the units over which policies are evaluated and datasets are organized. Why it matters for physical AI: Resets are cheap in simulation but expensive on hardware, where a human often must restage the scene, making episode design and reset-free learning active concerns for scaling real-world data collection. ### Episode Format Category: Data & Benchmarks URL: https://almond.bot/glossary/episode-format Episode Format is the on-disk schema used to store robot interaction episodes, specifying how time-aligned streams of camera frames, proprioception, actions, language annotations, and metadata are serialized. Common choices include RLDS built on TFRecords, used by Open X-Embodiment, HDF5 layouts popularized by robomimic, and the LeRobot format, which stores video-compressed frames with tabular data. Formats trade random-access speed, compression, and streaming behavior against tooling compatibility. Why it matters for physical AI: Format fragmentation is a real tax on the field: pooling datasets across labs and training on mixtures requires costly conversion, so converging on standard episode schemas accelerates every downstream effort. ### Epoch Category: Robot Learning URL: https://almond.bot/glossary/epoch Epoch is one complete pass of a training algorithm over the entire training dataset. Training typically spans many epochs with the data reshuffled each pass, and validation metrics tracked per epoch guide learning-rate schedules, checkpoint selection, and early stopping. For very large or streaming datasets, training is often measured in gradient steps or tokens instead, with less than one nominal epoch completed. Why it matters for physical AI: On small demonstration datasets policies can overfit within tens of epochs while action accuracy keeps improving, so epoch-level checkpointing paired with real rollout evaluation, not just validation loss, guides model selection. ### Equivariant Neural Network Category: Robot Learning URL: https://almond.bot/glossary/equivariant-neural-network Equivariant Neural Network is a network architecturally constrained so that transformations of its input, such as rotations or translations, produce correspondingly transformed outputs, encoding symmetry as an inductive bias rather than learning it from data. Group-equivariant CNNs (Cohen and Welling, 2016) generalized translation equivariance to rotations, and SE(3)-equivariant architectures operate on 3D point clouds and vector fields. In manipulation, equivariant policies and grasp detectors exploit the rotational symmetry of tabletop tasks. Why it matters for physical AI: Symmetry constraints can substitute for data: equivariant grasp and policy models reach strong performance from far fewer demonstrations, a meaningful lever when robot data collection dominates cost. ### EtherCAT Category: Software & Middleware URL: https://almond.bot/glossary/ethercat EtherCAT is a real-time industrial Ethernet fieldbus, introduced by Beckhoff in 2003 and standardized in IEC 61158, in which slave devices process frames on the fly as they pass through a daisy-chained ring. This yields microsecond-level cycle times and sub-microsecond synchronization via distributed clocks across hundreds of nodes. It is a dominant bus for coordinating servo drives, encoders, and I/O in robots, including many humanoid and quadruped platforms. Why it matters for physical AI: Kilohertz whole-body control across dozens of joints requires deterministic, synchronized communication that standard networking cannot provide, making buses like EtherCAT the invisible backbone of dynamic robot hardware. ### Euler Angles Category: Math & Kinematics URL: https://almond.bot/glossary/euler-angles Euler Angles are a parameterization of 3D orientation as three successive rotations about specified axes, with conventions such as ZYX yaw-pitch-roll common in robotics and aerospace. They are minimal and human-readable but suffer from gimbal lock, a singularity where two rotation axes align and a degree of freedom is lost, and from convention ambiguity across tools. Quaternions and rotation matrices are preferred for computation, with Euler angles retained for interfaces and displays. Why it matters for physical AI: Orientation representation choices propagate into learned systems: regressing Euler angles introduces discontinuities and singularities that degrade pose estimation and policy outputs, motivating quaternion or 6D rotation parameterizations. ### Euler-Lagrange Equation Category: Math & Kinematics URL: https://almond.bot/glossary/euler-lagrange-equation Euler-Lagrange Equation is the differential equation of motion obtained from a system's Lagrangian, the difference between kinetic and potential energy, via the principle of stationary action. Applied to a manipulator with joint coordinates, it yields the standard robot dynamics form involving the mass matrix, Coriolis and centrifugal terms, and gravity vector. This formulation is the basis for computed-torque control, inverse dynamics, and the dynamics engines inside physics simulators. Why it matters for physical AI: The manipulator equation derived this way is what simulators integrate and model-based controllers invert, so its structure quietly defines both the training environments and the control layers beneath learned policies. ### Eureka Category: Robot Learning URL: https://almond.bot/glossary/eureka Eureka is a 2023 NVIDIA method that uses a large language model, originally GPT-4, to automatically write and evolve reward functions for reinforcement learning. Given a task description and the simulation environment's source code, the LLM generates candidate reward code, receives training statistics as feedback, and iteratively refines its proposals. Eureka-designed rewards outperformed human-engineered ones on most benchmark tasks and enabled simulated dexterous pen spinning. Why it matters for physical AI: Reward engineering is a chief bottleneck of applied reinforcement learning, and code-writing language models turn it into an automated search, foreshadowing self-improving skill acquisition pipelines. ### Event Camera Category: Hardware URL: https://almond.bot/glossary/event-camera Event Camera is a neuromorphic vision sensor whose pixels asynchronously and independently report logarithmic brightness changes as timestamped events, rather than capturing full frames at a fixed rate. Commercialized as dynamic vision sensors by companies including iniVation and Prophesee, event cameras offer microsecond temporal resolution, roughly 120 dB dynamic range, and low power, at the cost of requiring fundamentally different processing algorithms. Why it matters for physical AI: High-speed motion and harsh lighting defeat conventional cameras exactly where dynamic robots operate, and event sensing has enabled low-latency drone obstacle avoidance and tracking beyond frame-based limits. ### Event-Triggered Control Category: Control URL: https://almond.bot/glossary/event-triggered-control Event-Triggered Control is a control paradigm in which measurements are sampled and control signals updated only when a state-dependent condition fires, such as the error exceeding a threshold, rather than at a fixed clock rate. Analysis guarantees stability while excluding Zeno behavior of infinitely frequent triggering. It substantially reduces communication and computation compared to periodic control, especially over shared or wireless networks. Why it matters for physical AI: Battery-powered fleets and networked robot systems face hard bandwidth and energy budgets, and aperiodic control offers a principled way to spend communication only when the physics demands it. ### Exoskeleton Category: Applications URL: https://almond.bot/glossary/exoskeleton Exoskeleton is a wearable robotic structure that operates in parallel with the human body to augment strength, assist movement, or restore mobility. Systems range from powered medical exoskeletons enabling gait rehabilitation and paraplegic walking, such as those from Ekso Bionics and ReWalk, to passive and powered occupational devices reducing shoulder and back load in industrial work. Design centers on human-robot kinematic alignment, transparent low-impedance actuation, and intent detection from motion or EMG signals. Why it matters for physical AI: Exoskeletons are physical AI worn on the body, and their intent-inference and compliant-actuation advances feed humanoid design, while instrumented suits double as natural-motion capture devices for manipulation data. ### Expert Demonstrations Category: Robot Learning URL: https://almond.bot/glossary/expert-demonstrations Expert Demonstrations are task executions performed by a proficient demonstrator, human or scripted controller, recorded as the supervision data for imitation learning. Expertise implies near-optimal, consistent behavior; in practice demonstrations vary in strategy, smoothness, and error-recovery content, and this heterogeneity strongly affects what behavior cloning learns. Offline reinforcement learning methods explicitly relax the expertise assumption to learn from mixed-quality data. Why it matters for physical AI: Policy quality is bounded by demonstration quality under pure imitation, so operator training, consistency audits, and quality-weighted sampling are as consequential as model architecture in production pipelines. ### Explicit Policy Category: Robot Learning URL: https://almond.bot/glossary/explicit-policy Explicit Policy is a policy represented as a direct feed-forward mapping from observations to actions, computed in a single network pass, in contrast to implicit policies that define actions indirectly as the minimizer of a learned energy function or the endpoint of an iterative sampling process. Explicit regression policies are fast and simple but represent only unimodal action distributions, averaging across conflicting demonstrations; the distinction was sharpened by the Implicit Behavioral Cloning work of 2021. Why it matters for physical AI: The explicit-implicit trade-off, one fast forward pass versus expressive multimodal generation, drives current policy architecture choices and the push to distill diffusion policies into fast explicit ones for high-rate control. ### Exploration Category: Robot Learning URL: https://almond.bot/glossary/exploration Exploration is the process by which a reinforcement learning agent tries actions of uncertain value to gather information about its environment, rather than exploiting current knowledge. Techniques range from action-space noise and epsilon-greedy selection to directed methods using optimism, count-based novelty, curiosity signals such as Random Network Distillation, and maximum-entropy objectives as in Soft Actor-Critic. Sparse-reward tasks make effective exploration the dominant difficulty. Why it matters for physical AI: Unstructured exploration is slow and hazardous on physical hardware, which is why real-robot pipelines lean on demonstrations, simulation pretraining, and safe exploration constraints rather than trial-and-error from scratch. ### Exploration-Exploitation Tradeoff Category: Robot Learning URL: https://almond.bot/glossary/exploration-exploitation-tradeoff Exploration-Exploitation Tradeoff is the fundamental dilemma of sequential decision-making: an agent must balance exploiting the best-known action against exploring alternatives that might prove better, since every exploratory choice forgoes known reward while every exploitative choice forgoes information. Formalized in the multi-armed bandit literature, principled resolutions include upper confidence bounds, Thompson sampling, and entropy-regularized objectives. Why it matters for physical AI: Deployed robots face the tradeoff operationally: a fleet can keep executing its proven behavior or trial updated policies that might fail, making bandit-style evaluation strategies relevant to continuous fleet improvement. ### Exteroception Category: Perception URL: https://almond.bot/glossary/exteroception Exteroception is sensing directed at the external environment, encompassing cameras, LiDAR, radar, microphones, and tactile arrays, in contrast to proprioception, which senses the robot's own internal state such as joint positions and body orientation. The distinction structures robot state estimation and policy design: exteroceptive streams are high-dimensional and semantically rich but subject to occlusion and lighting, while proprioceptive streams are low-dimensional, fast, and reliable. Why it matters for physical AI: How policies fuse exteroception with proprioception defines their robustness envelope, exemplified by legged controllers that walk proprioceptively when vision fails but anticipate terrain when it is available. ### Extrinsic Calibration Category: Perception URL: https://almond.bot/glossary/extrinsic-calibration Extrinsic Calibration is the estimation of the rigid-body transform, rotation and translation, between the coordinate frames of two devices, such as camera-to-robot-base, camera-to-camera, or LiDAR-to-IMU. Camera-to-arm calibration is the classical hand-eye problem, formulated as AX = XB and solved from paired motions while observing a target such as a checkerboard or AprilTag board. Errors in extrinsics translate directly into Cartesian task errors. Why it matters for physical AI: Visuomotor policies silently bake camera extrinsics into their weights, so a bumped camera can break a trained policy; calibration accuracy and recalibration workflows are recurring operational concerns for deployed manipulation. ### Extrinsic Contact Category: Manipulation URL: https://almond.bot/glossary/extrinsic-contact Extrinsic Contact is contact between a grasped object and the external environment, as distinct from the intrinsic contacts between the gripper and the object itself. Manipulation research uses extrinsic contacts deliberately, pressing a held part against a surface, edge, or fixture to pivot, slide, or reorient it, a strategy termed extrinsic dexterity by Dafle and Rodriguez. Estimating and controlling these contacts typically combines force-torque sensing, tactile feedback, and contact-mode reasoning. Why it matters for physical AI: Simple parallel-jaw grippers can achieve hand-like reorientation by borrowing dexterity from the environment, and insertion and assembly tasks hinge on regulating extrinsic contact forces, a frontier for tactile-aware learned policies. ## Glossary: F ### Factor Graph Category: Navigation & SLAM URL: https://almond.bot/glossary/factor-graph A factor graph is a bipartite probabilistic graphical model that connects variable nodes, such as robot poses and landmarks, to factor nodes encoding measurement constraints between them. Factor graphs are the dominant formulation for modern SLAM and state estimation, solved via nonlinear least-squares smoothing with libraries like GTSAM and incremental solvers such as iSAM2. They generalize pose graphs by supporting arbitrary sensor factors, including IMU preintegration, GPS, and visual landmarks. Why it matters for physical AI: Robots operating in unstructured environments must fuse many noisy sensors into a consistent state estimate; factor graphs provide the scalable, principled inference backbone for localization, mapping, and calibration on deployed systems. ### Failure Recovery Category: Robot Learning URL: https://almond.bot/glossary/failure-recovery Failure recovery is a robot's ability to detect that a task execution has gone wrong and take corrective action to return to a state from which the task can be completed. Approaches range from hand-coded retry logic and finite state machines to learned recovery policies trained on failure data, and vision-language models that verify subgoal completion and replan. Recovery data is often deliberately collected, since expert demonstrations rarely contain mistakes. Why it matters for physical AI: Long-horizon autonomy is limited less by peak success rate than by compounding errors; policies that recognize and correct their own failures are what separate lab demos from robots that run unattended for hours. ### Fall Recovery Category: Locomotion URL: https://almond.bot/glossary/fall-recovery Fall recovery is the behavior by which a legged robot returns to a nominal standing posture after falling, typically by sequencing contacts with its limbs and torso to right itself. Classical approaches scripted keyframe getup motions, while modern quadrupeds and humanoids increasingly use reinforcement learning policies trained in simulation to produce robust, contact-rich recovery from arbitrary fallen configurations. Why it matters for physical AI: Falls are inevitable for legged robots in the real world, and a machine that must be manually reset cannot be deployed at scale; reliable self-righting directly determines operational uptime and safety. ### FANUC Category: Hardware URL: https://almond.bot/glossary/fanuc FANUC is a Japanese industrial automation company and one of the world's largest manufacturers of industrial robots, CNC systems, and factory automation equipment. Founded in 1956 as a Fujitsu subsidiary and headquartered near Mount Fuji, FANUC is known for its yellow six-axis arms, delta robots, and collaborative robot lines, with an installed base spanning automotive, electronics, and general manufacturing worldwide. Why it matters for physical AI: Incumbent industrial robot fleets like FANUC's define the baseline of reliability and cost that learning-based systems must beat, and their factories are a primary target market for retrofitting foundation-model-driven skills. ### FAST (Frequency-space Action Sequence Tokenization) Category: Foundation Models URL: https://almond.bot/glossary/fast FAST is an action tokenization scheme, introduced by Physical Intelligence in 2025, that compresses continuous robot action chunks using a discrete cosine transform followed by byte-pair encoding. By decorrelating action sequences in frequency space before quantization, FAST yields far shorter token sequences than per-dimension binning, enabling efficient autoregressive training of vision-language-action models such as pi0-FAST on high-frequency dexterous data. Why it matters for physical AI: How actions are represented as tokens directly bounds a VLA model's control frequency, training efficiency, and dexterity; compression schemes like FAST make autoregressive transformers practical for fine-grained real-robot control. ### Feature Extraction Category: Perception URL: https://almond.bot/glossary/feature-extraction Feature extraction is the process of transforming raw sensor data, such as images or point clouds, into compact representations that capture information useful for downstream tasks. Classical pipelines used hand-engineered descriptors like SIFT, ORB, and FPFH, while modern systems learn features with convolutional networks or vision transformers, often pretrained at scale (for example DINOv2 or CLIP encoders) and reused across detection, pose estimation, and policy learning. Why it matters for physical AI: Robot policies are only as good as the representations they consume; strong pretrained visual features are a key ingredient that lets manipulation and navigation models generalize from limited robot data. ### Feature Matching Category: Perception URL: https://almond.bot/glossary/feature-matching Feature matching is the task of establishing correspondences between local features detected in two or more images or point clouds, typically by comparing descriptors and enforcing geometric consistency. It underpins visual odometry, SLAM loop closure, structure from motion, and object pose estimation. Learned matchers such as SuperGlue and LoFTR have largely surpassed classical nearest-neighbor matching with ratio tests on difficult viewpoint and illumination changes. Why it matters for physical AI: Reliable correspondences are the raw material of spatial understanding; robust matching under real-world lighting, motion blur, and texture-poor scenes determines whether a robot can localize and reconstruct its workspace. ### Feedback Linearization Category: Control URL: https://almond.bot/glossary/feedback-linearization Feedback linearization is a nonlinear control technique that algebraically cancels a system's nonlinear dynamics through state feedback and a coordinate transformation, rendering the closed-loop input-output behavior exactly linear so that standard linear controllers can be applied. For fully actuated manipulators it appears as computed torque control, which inverts the rigid-body dynamics model. The method requires an accurate dynamics model and full state measurement, and degrades under model mismatch. Why it matters for physical AI: Model-based cancellation of dynamics remains a workhorse for precise arm tracking, and its sensitivity to modeling error motivates hybrid schemes where learned residual models correct the nominal dynamics. ### Few-Shot Imitation Category: Robot Learning URL: https://almond.bot/glossary/few-shot-imitation Few-shot imitation is the ability of a robot to acquire a new task from only a handful of demonstrations, often just one, rather than the hundreds typically needed for behavior cloning. Approaches include meta-learning across task distributions, as in one-shot imitation learning by Duan et al. (2017), retrieval-augmented policies, and prompting large pretrained robot models with demonstrations as context. Why it matters for physical AI: Demonstration collection is the dominant cost in scaling robot skills; models that adapt from a few examples make deploying new tasks on-site economically viable rather than a data-engineering project. ### Fieldbus Category: Software & Middleware URL: https://almond.bot/glossary/fieldbus A fieldbus is an industrial digital communication network that connects controllers to sensors, drives, and actuators over a shared bus, replacing point-to-point analog wiring. Common variants in robotics include EtherCAT, CANopen, PROFINET, and EtherNet/IP, which differ in topology, determinism, and cycle rates. Real-time Ethernet fieldbuses like EtherCAT achieve sub-millisecond, low-jitter cycles suitable for closing torque loops across dozens of joint drives. Why it matters for physical AI: High-rate, deterministic buses are what allow a learned policy running on a compute module to command whole-body torque targets reliably; bus jitter and latency place hard limits on achievable control bandwidth. ### Figure 01 Category: Hardware URL: https://almond.bot/glossary/figure-01 Figure 01 is the first general-purpose humanoid robot developed by Figure AI, unveiled in 2023 as an electric, roughly human-scale bipedal platform aimed at commercial labor tasks. It gained attention through a BMW manufacturing partnership and a 2024 demonstration of speech-driven manipulation using OpenAI-powered reasoning. Figure subsequently iterated to Figure 02 and 03 and introduced Helix, an in-house vision-language-action model controlling the robot's upper body. Why it matters for physical AI: Commercial humanoid programs like Figure's are proving grounds for end-to-end neural control at industrial reliability, and their hardware iterations show how quickly the platform layer beneath robot foundation models is evolving. ### FiLM Conditioning Category: Robot Learning URL: https://almond.bot/glossary/film-conditioning FiLM (Feature-wise Linear Modulation) conditioning is a neural network technique, introduced by Perez et al. (2018), in which a conditioning signal generates per-channel scale and shift parameters that modulate intermediate feature maps of another network. In robotics it is widely used to inject language embeddings or goal information into visual encoders, as in BC-Z and RT-1, giving a cheap, effective way to make a single policy task-conditional. Why it matters for physical AI: Multi-task robot policies need their perception to be steered by instructions; lightweight conditioning mechanisms like FiLM let one network serve many tasks without duplicating encoders or retraining per skill. ### Fine-Tuning Category: Foundation Models URL: https://almond.bot/glossary/fine-tuning Fine-tuning is the process of continuing to train a pretrained model on a smaller, task- or domain-specific dataset so it specializes while retaining broadly useful representations. In robotics, vision-language-action models pretrained on large cross-embodiment corpora such as Open X-Embodiment are fine-tuned on demonstrations from a target robot and task, with variants including full-parameter updates, parameter-efficient methods like LoRA, and post-training with reinforcement learning. Why it matters for physical AI: Pretraining supplies general visuomotor priors, but deployments succeed or fail on adaptation; fine-tuning recipes determine how many demonstrations, how much compute, and how much forgetting a new skill costs. ### Finger Gaiting Category: Manipulation URL: https://almond.bot/glossary/finger-gaiting Finger gaiting is an in-hand manipulation strategy in which fingers are sequentially lifted, repositioned, and re-engaged on an object so it can be reoriented continuously beyond the workspace of any fixed grasp. Because contacts are broken and remade, the hand must maintain grasp stability with the remaining fingers throughout. Reinforcement learning systems such as OpenAI's Rubik's cube work demonstrated emergent finger gaiting on high-DoF hands. Why it matters for physical AI: Dexterity beyond pick-and-place requires reorienting objects within the hand; finger gaiting is a core capability separating simple grippers from human-level manipulation and a benchmark for contact-rich policy learning. ### Fingertip Sensor Category: Hardware URL: https://almond.bot/glossary/fingertip-sensor A fingertip sensor is a tactile sensing element mounted at the distal end of a robotic finger to measure contact properties such as normal and shear force, contact location, vibration, and local geometry. Technologies include capacitive and piezoresistive arrays, barometric domes, and vision-based elastomer sensors like GelSight and DIGIT, which image the deformation of a gel pad to recover high-resolution contact maps. Why it matters for physical AI: Vision alone cannot resolve slip, contact force, or occluded geometry at the moment of grasp; fingertip sensing supplies the feedback signal that makes learned manipulation robust to the last millimeter of uncertainty. ### Finite State Machine (FSM) Category: Software & Middleware URL: https://almond.bot/glossary/finite-state-machine A finite state machine is a computational model consisting of a finite set of discrete states, transitions triggered by events or conditions, and actions associated with states or transitions. In robotics, FSMs structure task-level logic and mode switching, such as locomotion phase machines in legged control or grasp-transport-place sequences in manipulation. Hierarchical state machines and behavior trees extend the pattern for more complex, reusable task logic. Why it matters for physical AI: Even highly learned stacks need auditable supervisory logic for mode switching, safety interlocks, and recovery sequencing; state machines remain the standard glue between neural policies and dependable system behavior. ### Fisheye Camera Category: Perception URL: https://almond.bot/glossary/fisheye-camera A fisheye camera is a camera fitted with an ultra-wide-angle lens, typically covering 180 degrees or more, that produces strongly distorted images following non-pinhole projection models such as equidistant or double-sphere. Fisheye optics trade angular resolution for coverage, and their images require dedicated calibration and undistortion models. They are common in robot navigation, visual-inertial odometry rigs, and surround-view systems where field of view matters more than metric detail. Why it matters for physical AI: Wide field of view reduces blind spots for navigation and whole-scene manipulation context with fewer cameras, but the distortion challenges standard vision models, making fisheye handling a practical concern in camera stack design. ### Fixed-Base Manipulator Category: Hardware URL: https://almond.bot/glossary/fixed-base-manipulator A fixed-base manipulator is a robotic arm rigidly mounted to a stationary structure such as a table, pedestal, or cell floor, in contrast to mobile manipulators whose base can move. The fixed base provides a stable inertial reference, simplifying kinematics, calibration, and control, at the cost of a workspace bounded by the arm's reach. Most industrial arms and research platforms like the Franka Panda are used in this configuration. Why it matters for physical AI: Fixed-base arms are where most manipulation datasets and benchmarks originate, so policies must eventually bridge the gap from this controlled setting to mobile bases with shifting viewpoints and unmodeled base motion. ### Flatness Category: Control URL: https://almond.bot/glossary/flatness Flatness, or differential flatness, is a property of a dynamical system whereby all states and inputs can be expressed algebraically in terms of a set of flat outputs and finitely many of their time derivatives. Flat systems admit straightforward trajectory generation: one plans a smooth curve in flat output space and recovers the full state and input trajectories analytically. The quadrotor with its position and yaw as flat outputs is the canonical robotics example. Why it matters for physical AI: Flatness turns aggressive trajectory planning into curve fitting, enabling agile drone flight and feedforward-rich control; it exemplifies how exploiting structure yields guarantees that purely learned planners still struggle to match. ### Fleet Learning Category: Robot Learning URL: https://almond.bot/glossary/fleet-learning Fleet learning is the practice of improving robot capabilities by aggregating experience from many deployed robots, retraining shared models on the pooled data, and redistributing updated policies across the fleet. The paradigm, popularized in autonomous driving, extends to manipulation fleets where interventions, failures, and successes from the field become training data. Infrastructure concerns include data triage, automatic labeling, evaluation gating, and staged rollout of new policies. Why it matters for physical AI: A single robot gathers experience slowly, but a fleet compounds it; the data flywheel from deployed units is the central economic argument for scaling robot foundation models in the real world. ### Flexible Joint Category: Hardware URL: https://almond.bot/glossary/flexible-joint A flexible joint is a robot joint exhibiting non-negligible elasticity between the actuator and the driven link, arising from compliant transmissions such as harmonic drives, belts, or deliberately introduced springs as in series elastic actuators. The elasticity adds unactuated dynamics that can cause oscillation and complicates control, motivating techniques like full-state feedback on motor and link positions, or singular perturbation approaches. It also provides beneficial shock tolerance and force sensing. Why it matters for physical AI: Compliance protects hardware and humans during unexpected contact and enables torque estimation from spring deflection, but the added dynamics must be modeled or learned for precise, high-bandwidth manipulation. ### Flight Phase Category: Locomotion URL: https://almond.bot/glossary/flight-phase A flight phase is an interval during legged locomotion in which no part of the robot contacts the ground, occurring in dynamic gaits such as running, trotting with suspension, pronking, and jumping. During flight the center of mass follows ballistic dynamics that no ground force can alter, so the controller can only reorient limbs and body angular momentum to prepare touchdown conditions. Flight phases make gait timing and landing control critical. Why it matters for physical AI: Dynamic gaits with aerial phases are essential for speed and obstacle negotiation, and their underactuated ballistic dynamics stress-test learned locomotion policies far beyond quasi-static walking. ### FLIR Camera Category: Hardware URL: https://almond.bot/glossary/flir-camera A FLIR camera is a thermal imaging camera that forms images from long-wave infrared radiation emitted by objects, named after FLIR Systems (now part of Teledyne), the dominant manufacturer whose brand became shorthand for the sensor class. Thermal cameras perceive temperature contrast rather than reflected light, operating in darkness, smoke, and fog. In robotics they serve inspection, firefighting, search and rescue, and human detection tasks; Teledyne FLIR also sells widely used machine-vision RGB cameras under the Blackfly line. Why it matters for physical AI: Sensing modalities beyond visible light extend where robots can operate; thermal imaging gives reliable human and hotspot detection for safety and inspection missions that RGB-only perception stacks cannot cover. ### Flow Matching Category: Robot Learning URL: https://almond.bot/glossary/flow-matching Flow matching is a generative modeling framework, formalized by Lipman et al. (2023), that trains a neural network to regress a time-dependent velocity field transporting samples from a simple noise distribution to the data distribution along continuous probability paths. It offers a simulation-free training objective closely related to diffusion models but typically permits faster sampling with fewer integration steps, making it attractive for generating continuous, multimodal outputs. Why it matters for physical AI: Robot action distributions are continuous and multimodal, and inference must run at control rate; flow matching delivers diffusion-quality action generation with fewer denoising steps, easing the latency constraints of real-time policies. ### Flow Matching Policy Category: Robot Learning URL: https://almond.bot/glossary/flow-matching-policy A flow matching policy is a robot control policy that generates action chunks by integrating a learned velocity field from noise to actions, conditioned on observations such as images, proprioception, and language. Physical Intelligence's pi0 is the best-known example, attaching a flow matching action expert to a pretrained vision-language model to produce continuous 50 Hz actions for dexterous tasks like laundry folding. Why it matters for physical AI: Continuous action heads determine how precisely and quickly a foundation model can move a real robot; flow matching heads have become a leading design for high-frequency dexterous control in modern VLAs. ### FMCG Robot Category: Applications URL: https://almond.bot/glossary/fmcg-robot An FMCG robot is a robotic system deployed in the fast-moving consumer goods sector, handling high-throughput tasks such as primary and secondary packaging, case packing, palletizing, and order fulfillment for products like food, beverages, and toiletries. Typical hardware includes delta robots for high-speed picking, articulated arms for palletizing, and increasingly AI-driven piece-picking cells that cope with the sector's enormous SKU variety and frequent product changeovers. Why it matters for physical AI: FMCG operations combine extreme cycle-rate demands with constant product variation, making the sector a proving ground for vision-guided picking policies that must generalize across thousands of SKUs without per-item engineering. ### Food Handling Robot Category: Applications URL: https://almond.bot/glossary/food-handling-robot A food handling robot is a robotic system designed to grasp, portion, assemble, or package food items, operating under hygiene regulations that demand washdown-rated, food-safe materials and easy-clean designs. Because food is deformable, slippery, variable, and delicate, these systems favor soft grippers, suction, and compliant control, with vision-based pose estimation handling natural variation. Applications span meal assembly, meat and produce processing, bakery lines, and quick-service restaurant automation. Why it matters for physical AI: Deformable, variable, and fragile items defeat rigid automation assumptions, so food handling is a demanding benchmark for learned grasping and force control, and a large labor market awaiting capable general-purpose manipulation. ### Foot Clearance Category: Locomotion URL: https://almond.bot/glossary/foot-clearance Foot clearance is the height by which a legged robot's swing foot passes above the ground or obstacles during the swing phase of a gait. Insufficient clearance causes trips and scuffs on uneven terrain, while excessive clearance wastes energy and slows the gait. Controllers regulate clearance through swing trajectory design, and learned locomotion policies are often shaped with clearance rewards or terrain-aware foot placement to traverse rubble and stairs. Why it matters for physical AI: Trips on unseen ground are a leading failure mode for legged deployment; clearance behavior encodes how well a policy has internalized terrain geometry versus merely memorizing flat-ground gaits. ### Footstep Planning Category: Locomotion URL: https://almond.bot/glossary/footstep-planning Footstep planning is the problem of selecting a sequence of foot placement positions, orientations, and timings that carry a legged robot to a goal while respecting kinematic reach, balance constraints, and terrain validity. Methods include graph search over discrete step sets, mixed-integer and nonlinear optimization over stepping stones, and learned policies that propose footholds from perception. It is typically layered above whole-body and below global path planning. Why it matters for physical AI: Where a robot may step is the fundamental contract between perception and legged control; robust foothold selection on stairs, gaps, and rubble is what turns locomotion policies into navigation capability. ### Force Closure Category: Manipulation URL: https://almond.bot/glossary/force-closure Force closure is a grasp property in which the contact forces that the fingers can apply, accounting for friction, span the entire wrench space, allowing the grasp to resist any external force and torque applied to the object. A force-closure grasp can hold an object against arbitrary disturbances provided contact forces remain within their friction cones. It is a classical criterion in grasp analysis and a standard filter in grasp planning pipelines. Why it matters for physical AI: Grasp stability guarantees still anchor how planners and learned grasp detectors are evaluated; force closure provides the physics-based ground truth against which data-driven grasp quality predictions are validated. ### Force Control Category: Control URL: https://almond.bot/glossary/force-control Force control is a family of robot control strategies that regulate the forces and torques a robot exerts on its environment, rather than or in addition to its position. Variants include direct force control with force-torque feedback, hybrid position/force control that partitions task axes, and indirect methods like impedance and admittance control that shape the relationship between motion and force. It is essential for contact-rich tasks such as insertion, polishing, and assembly. Why it matters for physical AI: Contact-rich manipulation fails under pure position control because millimeter errors produce huge forces; force-aware control is the substrate that lets learned policies act safely and effectively through contact. ### Force-Torque Sensor (F/T Sensor) Category: Hardware URL: https://almond.bot/glossary/force-torque-sensor A force-torque sensor is a transducer, most commonly a six-axis strain-gauge or capacitive device mounted at a robot's wrist, that measures the three force and three torque components transmitted through it. F/T sensors provide the feedback for force control, contact detection, insertion tasks, and haptic teleoperation, and their signals require bias calibration and gravity compensation of the attached tool mass. Joint-level torque sensing serves a similar role distributed through the arm. Why it matters for physical AI: Wrench measurements give policies a direct window into contact that vision cannot provide, enabling assembly-grade precision, safe human interaction, and rich data for training contact-aware manipulation models. ### Form Closure Category: Manipulation URL: https://almond.bot/glossary/form-closure Form closure is a grasp or fixture property in which an object's mobility is completely constrained by the geometry of frictionless contacts alone, so no motion is possible regardless of applied wrenches. Because it assumes no friction, form closure is a stronger condition than force closure and generally requires more contacts, at least four in the plane and seven in space for a rigid body. It underlies fixture and caging design. Why it matters for physical AI: Geometry-only immobilization offers robustness that does not depend on uncertain friction estimates, informing caging grasps and fixture strategies that make manipulation reliable when surface properties are unknown. ### Forward Kinematics (FK) Category: Math & Kinematics URL: https://almond.bot/glossary/forward-kinematics Forward kinematics is the computation of a robot's end-effector pose, or the poses of any of its links, from its joint positions using the kinematic model of the chain. It is evaluated by composing joint transforms, classically parameterized with Denavit-Hartenberg conventions or the product of exponentials formulation. FK is deterministic and unique for serial chains, in contrast to inverse kinematics, which may admit multiple or no solutions. Why it matters for physical AI: Every layer of a robot stack, from Cartesian teleoperation to action-space design for learned policies, relies on fast, differentiable FK; it is the basic map between the joint space where motors live and the task space where goals live. ### Foundation Model Category: Foundation Models URL: https://almond.bot/glossary/foundation-model A foundation model is a large neural network pretrained on broad data at scale such that it can be adapted, via fine-tuning or prompting, to a wide range of downstream tasks. The term was coined by Stanford researchers in 2021 in the wake of models like GPT-3 and CLIP. In robotics, the concept extends to robot foundation models such as RT-2, pi0, and GR00T N1, trained on large multi-embodiment datasets like Open X-Embodiment to produce generalist visuomotor policies. Why it matters for physical AI: The bet behind modern physical AI is that broad pretraining transfers to embodied control the way it did to language and vision, replacing task-specific engineering with generalist models adapted per deployment. ### Fourier GR-1 Category: Hardware URL: https://almond.bot/glossary/fourier-gr-1 The Fourier GR-1 is a full-size humanoid robot developed by Shanghai-based Fourier Intelligence (now Fourier), announced in 2023 as one of the first commercially available general-purpose humanoid platforms. The GR-1 stands about 1.65 m tall with roughly 40 degrees of freedom driven by the company's proprietary actuators, and has been adopted by research groups as a real-robot target for humanoid whole-body control and learning experiments; Fourier followed it with the GR-2. Why it matters for physical AI: Commercially purchasable humanoids like the GR-1 lower the barrier for academic and industrial teams to test locomotion and manipulation policies on real full-size hardware rather than simulation alone. ### Fourth Industrial Revolution Category: Applications URL: https://almond.bot/glossary/fourth-industrial-revolution The Fourth Industrial Revolution, often called Industry 4.0, is the ongoing transformation of manufacturing and industry through cyber-physical systems, the industrial internet of things, cloud computing, big data, and AI, following mechanization, electrification, and digital automation. The term was popularized by Klaus Schwab and the World Economic Forum and by Germany's Industrie 4.0 initiative. Robotics features centrally through connected, flexible, and increasingly autonomous production systems. Why it matters for physical AI: The industrial appetite for flexible, data-driven automation frames the commercial context for general-purpose robots, where learning-based systems promise to automate the long tail of tasks fixed automation never reached. ### FPGA in Robotics Category: Hardware URL: https://almond.bot/glossary/fpga-in-robotics An FPGA (field-programmable gate array) is a reconfigurable integrated circuit whose logic fabric can be programmed to implement custom digital hardware, offering deterministic, low-latency, highly parallel processing. In robotics, FPGAs handle hard real-time workloads such as motor commutation and current loops in servo drives, encoder decoding, EtherCAT slave controllers, sensor interfacing, and low-latency vision preprocessing, often alongside CPUs in system-on-chip parts like the Xilinx Zynq family. Why it matters for physical AI: Microsecond-deterministic motor and sensor loops cannot run on general-purpose operating systems; FPGA layers provide the timing backbone that lets higher-level learned policies assume clean, high-rate actuation and sensing. ### FPN (Feature Pyramid Network) Category: Perception URL: https://almond.bot/glossary/fpn An FPN (Feature Pyramid Network) is a convolutional network architecture, introduced by Lin et al. (2017), that builds a multi-scale feature pyramid by combining a backbone's high-resolution shallow features with semantically strong deep features through top-down pathways and lateral connections. FPNs became a standard component of object detectors and segmentation models such as Faster R-CNN and Mask R-CNN, enabling robust recognition of objects across widely varying scales. Why it matters for physical AI: Robots encounter task-relevant objects at wildly different apparent sizes as distance varies; multi-scale feature architectures underpin the detectors robots use for grasp target localization and scene understanding. ### Franka Emika Panda Category: Hardware URL: https://almond.bot/glossary/franka-emika-panda The Franka Emika Panda is a 7-degree-of-freedom torque-controlled collaborative robot arm released in 2017, featuring joint torque sensors in every axis, a 3 kg payload, and a 1 kHz real-time control interface (libfranka and the Franka Control Interface). Its combination of compliant torque control, research-friendly software, and relatively low cost made it the de facto standard arm in manipulation research, appearing in countless learning papers and benchmarks. Why it matters for physical AI: Shared hardware platforms make robot learning results comparable and reproducible across labs; the Panda's ubiquity shaped the datasets, simulators, and action spaces much of today's manipulation research builds on. ### Franka FR3 Category: Hardware URL: https://almond.bot/glossary/franka-fr3 The Franka FR3, made by Franka Robotics, is the successor to the Franka Emika Panda: a 7-degree-of-freedom torque-controlled research arm with joint torque sensing, a 3 kg payload, and the same 1 kHz low-level control interface, updated for industrial certification and improved reliability. It preserves software compatibility with the libfranka ecosystem, allowing the large body of Panda-based research code and learning pipelines to carry over. Why it matters for physical AI: Continuity of a dominant research platform matters for data reuse; FR3 arms feature prominently in recent large-scale manipulation data collection efforts and keep the torque-controlled 7-DoF arm a lingua franca of policy learning. ### Friction Cone Category: Manipulation URL: https://almond.bot/glossary/friction-cone A friction cone is the set of contact forces that a frictional point contact can transmit without slipping, forming a cone around the surface normal whose half-angle is the arctangent of the friction coefficient under the Coulomb model. Grasp and locomotion planners constrain contact forces to lie within their friction cones, often using polyhedral approximations to keep the resulting optimization problems linear or conic. Why it matters for physical AI: Whether a foot or finger slips is governed by friction cone constraints, so they appear in nearly every contact-aware planner and controller, and violations are a dominant cause of real-world grasp and locomotion failures. ### Frontal Plane Control Category: Locomotion URL: https://almond.bot/glossary/frontal-plane-control Frontal plane control is the regulation of a legged robot's lateral balance and motion within the frontal (coronal) plane, the vertical plane dividing front from back, covering side-to-side center-of-mass sway, hip abduction and adduction, and lateral foot placement. Bipedal walking exhibits naturally unstable lateral rocking between support feet, so controllers stabilize it through lateral stepping strategies, ankle inversion torques, and step-width regulation, often analyzed separately from sagittal plane dynamics. Why it matters for physical AI: Lateral balance failures cause sideways falls that sagittal-plane analysis never predicts; explicit frontal plane treatment is essential for humanoids stepping in cluttered, narrow spaces around humans. ### Frustum Category: Perception URL: https://almond.bot/glossary/frustum A frustum is the truncated-pyramid volume of 3D space visible to a camera, bounded laterally by the projection of the image boundaries and in depth by near and far planes. In robot perception, frustums link 2D image evidence to 3D space: a 2D detection defines a frustum in which the object must lie, an idea exploited by Frustum PointNets for 3D detection, and frustum culling restricts processing to visible space. Why it matters for physical AI: Lifting 2D detections into 3D is a recurring need when robots act on camera-based perception; frustum reasoning provides the geometric bridge between image-space models and the metric world where grasps and motions are executed. ### Functional Safety Category: Safety & Standards URL: https://almond.bot/glossary/functional-safety Functional safety is the discipline of ensuring that a system's safety-related functions perform correctly in response to faults and hazardous conditions, reducing risk to acceptable levels through measures like redundant sensing, safe-state transitions, and monitored stop functions. It is governed by standards including IEC 61508, ISO 13849 for machinery control, and in robotics ISO 10218 and the ISO/TS 15066 guidance for collaborative operation, with safety integrity quantified by levels such as SIL and PL. Why it matters for physical AI: Certifying learned, non-deterministic policies against safety standards is an open challenge; today's deployments wrap neural controllers in certified supervisory layers, making functional safety architecture a gating factor for robots near people. ### Funnel Control Category: Control URL: https://almond.bot/glossary/funnel-control Funnel control is a model-free adaptive control methodology that guarantees a system's tracking error remains inside a prescribed, typically shrinking, time-varying boundary called the funnel, by increasing feedback gain as the error approaches the funnel wall. Introduced by Ilchmann, Ryan, and Sangwin (2002), it provides transient and steady-state performance guarantees without a plant model. Relatedly, funnels in motion planning denote verified regions of attraction composed sequentially, as in LQR-trees. Why it matters for physical AI: Prescribed-performance guarantees without accurate models suit robots whose dynamics are uncertain or learned, and funnel composition ideas inform how verified safety certificates can wrap around otherwise unverified learned controllers. ## Glossary: G ### GAIL (Generative Adversarial Imitation Learning) Category: Robot Learning URL: https://almond.bot/glossary/gail GAIL is an imitation learning algorithm, introduced by Ho and Ermon (2016), that trains a policy adversarially against a discriminator which tries to distinguish the policy's state-action occupancy from that of expert demonstrations. The discriminator's output serves as a reward signal for reinforcement learning, allowing the agent to match expert behavior without explicitly recovering a reward function and while gathering its own environment interactions. Why it matters for physical AI: Adversarial imitation can outperform behavior cloning when demonstrations are scarce, since the learner explores beyond demonstrated states; the idea shaped how reward signals are extracted from human data for robot training. ### Gain Scheduling Category: Control URL: https://almond.bot/glossary/gain-scheduling Gain scheduling is a control technique in which the parameters of a linear controller are varied as a function of measured operating conditions, allowing a family of locally tuned linear controllers to cover a nonlinear system's operating envelope. Gains are typically designed at a grid of operating points and interpolated online. It is standard practice in flight control and appears in robotics wherever dynamics change substantially, such as payload variation or configuration-dependent inertia. Why it matters for physical AI: Robot dynamics change dramatically with pose, payload, and contact state; scheduled or context-conditioned gains are a simple, certifiable precursor to the fully adaptive behavior that learned controllers aim to provide. ### Gait Library Category: Locomotion URL: https://almond.bot/glossary/gait-library A gait library is a precomputed collection of reference gaits or controllers, each optimized offline for a particular condition such as speed, terrain slope, or step length, from which a legged robot selects and interpolates online. The approach was prominent in hybrid zero dynamics work on bipeds like Cassie, where libraries of periodic orbits enabled robust velocity tracking. Learned policies sometimes distill or imitate such libraries. Why it matters for physical AI: Offline optimization can produce dynamically precise gaits that online computation cannot afford; libraries bridge model-based gait design and real-time control, and provide structured priors or training targets for learned locomotion. ### Gait Transition Category: Locomotion URL: https://almond.bot/glossary/gait-transition A gait transition is the change of a legged robot's locomotion pattern from one gait to another, such as a quadruped shifting from walk to trot to gallop, or a biped moving between standing, walking, and running. Transitions alter the contact sequence and timing, and must be executed without destabilizing the robot, typically triggered by commanded speed, energetic efficiency, or terrain. Learned locomotion policies can exhibit emergent, smooth transitions. Why it matters for physical AI: Real environments demand continuously varying speeds and terrains, so seamless gait switching without stumbles is a practical robustness marker for locomotion controllers deployed outside the lab. ### GAN (Generative Adversarial Network) Category: Robot Learning URL: https://almond.bot/glossary/gan A GAN is a generative model architecture, introduced by Goodfellow et al. (2014), in which a generator network learns to produce samples that a jointly trained discriminator network cannot distinguish from real data, framed as a two-player minimax game. GANs produced the first high-fidelity neural image synthesis and spawned variants like CycleGAN for unpaired image translation. In robotics they have been used for sim-to-real image adaptation, data augmentation, and adversarial imitation. Why it matters for physical AI: Although diffusion and flow models now dominate generation, adversarial training remains influential in robotics for domain adaptation and imitation, where matching distributions matters more than likelihood. ### GAN Domain Adaptation Category: Simulation URL: https://almond.bot/glossary/gan-domain-adaptation GAN domain adaptation is the use of generative adversarial networks to translate data between visual domains, most commonly rendering simulated images photorealistic or mapping real images into a canonical simulation-like appearance so that policies trained in one domain transfer to another. Notable robotics examples include GraspGAN for sim-to-real grasping and RCAN, which translated real images to a canonical sim domain for QT-Opt-style grasping, and RetinaGAN for detection-consistent translation. Why it matters for physical AI: The visual gap between simulation and reality is a principal obstacle to cheap simulated training; learned image translation offered an alternative to domain randomization and still informs modern sim-to-real pipelines. ### Gantry Robot Category: Hardware URL: https://almond.bot/glossary/gantry-robot A gantry robot is a Cartesian robot whose end effector moves along linear axes supported by an overhead bridge structure spanning the workspace, typically with X, Y, and Z rails in a rectangular frame. Gantries offer large, scalable work envelopes, high payloads, and stiff, repeatable positioning with simple kinematics, at the cost of floor space and limited dexterity. They are common in CNC machining, 3D printing, warehouse picking, and large-part handling. Why it matters for physical AI: For overhead coverage of large areas, gantries remain more economical than articulated arms, and they serve as stable camera and tool platforms in data collection rigs and automated labs feeding robot learning pipelines. ### Gato Category: Foundation Models URL: https://almond.bot/glossary/gato Gato is a generalist agent from DeepMind (2022): a single 1.2-billion-parameter transformer trained on around 600 tasks spanning Atari games, image captioning, dialogue, and real robot block stacking, with all modalities serialized into a common token sequence. Gato demonstrated that one set of weights could perform vision, language, and control tasks, though per-task performance trailed specialists. It is widely cited as a precursor to robot foundation models. Why it matters for physical AI: Gato provided early evidence that control can be folded into the same sequence-modeling recipe as language and vision, foreshadowing the token-based vision-language-action models now central to physical AI. ### Gauss-Newton Category: Math & Kinematics URL: https://almond.bot/glossary/gauss-newton Gauss-Newton is an iterative algorithm for nonlinear least-squares problems that approximates the Hessian using only first-order Jacobian information, solving a linearized normal-equation system at each step. It converges rapidly near solutions where residuals are small, and its damped variant, Levenberg-Marquardt, adds regularization for robustness. In robotics it is the standard solver core for bundle adjustment, factor-graph SLAM, camera calibration, and inverse kinematics. Why it matters for physical AI: Most geometric estimation a robot performs, from calibrating its cameras to building maps, reduces to nonlinear least squares; Gauss-Newton-family solvers are the numerical engine making these computations fast enough for online use. ### Gaussian Mixture Model (GMM) Category: Robot Learning URL: https://almond.bot/glossary/gaussian-mixture-model A Gaussian mixture model is a probabilistic model that represents a distribution as a weighted sum of Gaussian components, fitted with expectation-maximization. In robotics, GMMs have been widely used in learning from demonstration, where trajectories are encoded as mixtures and reproduced via Gaussian mixture regression, a line of work developed extensively by Calinon and colleagues. They also appear in clustering, multimodal state estimation, and as parametric action heads in some policies. Why it matters for physical AI: Demonstrations of the same task vary, and GMMs offered an early principled way to capture that multimodality; the same need now drives diffusion and flow-based action heads in modern policies. ### Gaussian Process (GP) Category: Robot Learning URL: https://almond.bot/glossary/gaussian-process A Gaussian process is a nonparametric Bayesian model that defines a distribution over functions, providing predictions with calibrated uncertainty from limited data via a kernel encoding smoothness assumptions. In robotics, GPs have modeled dynamics for data-efficient model-based reinforcement learning, most famously in PILCO, as well as terrain, residual dynamics errors, and Bayesian optimization objectives for controller tuning. Exact inference scales cubically with data, limiting GPs to modest datasets. Why it matters for physical AI: Sample efficiency and honest uncertainty are perpetual concerns when learning on physical hardware; GPs remain a benchmark for data-efficient model learning and a tool for safe exploration and controller auto-tuning. ### Gaussian Splatting Category: Perception URL: https://almond.bot/glossary/gaussian-splatting Gaussian splatting is a 3D scene representation and rendering technique, introduced as 3D Gaussian Splatting by Kerbl et al. (2023), that models a scene as millions of anisotropic 3D Gaussians with color and opacity, rendered by differentiable rasterization. It matches or exceeds NeRF visual quality while rendering orders of magnitude faster, and is fitted from posed images in minutes. Robotics applications include mapping, simulation asset creation, and photorealistic rendering for policy training. Why it matters for physical AI: Fast, editable, photoreal scene reconstructions narrow the sim-to-real visual gap and enable real-to-sim pipelines where robots train in digital twins of the exact environments they will operate in. ### Gazebo Category: Simulation URL: https://almond.bot/glossary/gazebo Gazebo is an open-source robotics simulator, originally developed in the early 2000s and long maintained alongside ROS by Open Robotics, providing rigid-body physics via engines like ODE, sensor simulation, and plugin-based robot models described in SDF or URDF. The classic version was succeeded by the modernized Gazebo (formerly Ignition) architecture. It remains a default choice for ROS-integrated testing of navigation, manipulation, and multi-robot systems. Why it matters for physical AI: Tight ROS integration makes Gazebo the standard rehearsal environment for full robot software stacks, though large-scale policy training has shifted toward GPU-parallel simulators optimized for reinforcement learning throughput. ### Gear Ratio Category: Hardware URL: https://almond.bot/glossary/gear-ratio Gear ratio is the ratio between input and output rotational speeds of a transmission, equal to the factor by which torque is multiplied and speed divided in an ideal reducer. Robot joints commonly use harmonic drives, planetary gearboxes, or cycloidal reducers with ratios from about 5:1 to over 100:1. Higher ratios raise torque density but amplify reflected inertia, friction, and backlash, degrading backdrivability and the fidelity of torque control through the transmission. Why it matters for physical AI: The choice between high-ratio precision and low-ratio quasi-direct-drive transmissions shapes whether a robot can feel and control contact, a hardware decision that directly constrains what contact-rich policies can achieve. ### GelSight Category: Hardware URL: https://almond.bot/glossary/gelsight GelSight is a vision-based tactile sensor technology, developed at MIT by Adelson and colleagues, in which a camera images the deformation of an illuminated elastomer gel pressed against a surface, recovering high-resolution contact geometry, texture, and force cues via photometric stereo. The approach yields per-pixel tactile maps far denser than taxel arrays, and inspired derivatives including GelSlim, DIGIT, and commercial GelSight sensors used on robot fingertips. Why it matters for physical AI: High-resolution touch turns contact from a binary event into a rich image-like observation that fits naturally into modern visuomotor learning, enabling policies that perceive slip, texture, and in-hand object pose. ### Gemini Robotics Category: Foundation Models URL: https://almond.bot/glossary/gemini-robotics Gemini Robotics is a family of vision-language-action models from Google DeepMind, introduced in 2025, that builds robot control on top of the Gemini multimodal foundation model. The initial release paired a VLA that outputs robot actions with Gemini Robotics-ER, an embodied-reasoning variant for spatial understanding, and demonstrated dexterous bimanual tasks and cross-embodiment operation; subsequent versions added on-device execution and agentic long-horizon behavior. Why it matters for physical AI: Frontier-lab entries like Gemini Robotics test whether internet-scale multimodal pretraining transfers decisive advantages to control, and their embodied-reasoning splits illustrate one architecture for coupling semantic planning with low-level action. ### Generalization (Robot Policy) Category: Robot Learning URL: https://almond.bot/glossary/generalization-robot-policy Generalization, in the context of robot policies, is the ability of a learned controller to succeed under conditions not seen during training, spanning axes such as novel object instances and categories, unseen backgrounds and lighting, new spatial arrangements, paraphrased language instructions, and entirely new embodiments. Benchmarks and studies, from RT-2's emergent semantic transfer to systematic evaluations like SIMPLER and COLOSSEUM, attempt to quantify these axes separately. Why it matters for physical AI: The economic promise of robot foundation models rests entirely on generalization, since per-deployment data collection is what makes traditional automation expensive; measuring it honestly is a central open problem. ### Generalized Coordinates Category: Math & Kinematics URL: https://almond.bot/glossary/generalized-coordinates Generalized coordinates are a minimal set of independent variables that completely specify the configuration of a mechanical system, such as the joint angles of a serial manipulator or the base pose plus joint angles of a floating-base robot. Lagrangian mechanics expresses dynamics directly in these coordinates, yielding the standard manipulator equation involving the mass matrix, Coriolis terms, and generalized forces. Their time derivatives are generalized velocities. Why it matters for physical AI: Every dynamics engine, whole-body controller, and trajectory optimizer operates on generalized coordinates; understanding them is prerequisite to reading the equations of motion that govern simulation and model-based control. ### Generative Robot Data Category: Data & Benchmarks URL: https://almond.bot/glossary/generative-robot-data Generative robot data is training data for robot policies produced partly or wholly by algorithms rather than direct human demonstration, including trajectory synthesis systems like MimicGen and DexMimicGen that expand a few human demonstrations into thousands of simulated variants, procedurally generated scenes and tasks as in RoboCasa, and image or video editing that augments real data. Quality control and physical plausibility are the central challenges. Why it matters for physical AI: Robot data is the scarcest resource in physical AI, orders of magnitude behind text corpora; generative multiplication of human effort is among the most credible paths to foundation-model-scale embodied datasets. ### Genesis Category: Simulation URL: https://almond.bot/glossary/genesis Genesis is an open-source physics simulation platform for robotics and embodied AI, released in late 2024 by a large academic-industry collaboration, combining a GPU-parallelized physics engine supporting rigid bodies, deformables, fluids, and coupling, with very high simulation throughput. The project also targets generative simulation, aiming to produce environments, tasks, and data from high-level prompts, and positions itself as a unified platform for training robot policies at scale. Why it matters for physical AI: Simulation speed and physics breadth set the ceiling on cheap embodied experience; massively parallel, differentiable-friendly platforms are core infrastructure in the race to train generalist policies without prohibitive real-robot data costs. ### Geometric Primitive Fitting Category: Perception URL: https://almond.bot/glossary/geometric-primitive-fitting Geometric primitive fitting is the estimation of parametric shapes, such as planes, spheres, cylinders, cones, and superquadrics, from 3D sensor data like point clouds, typically using RANSAC-style robust sampling, Hough transforms, or least-squares refinement. Fitted primitives compress noisy geometry into compact, semantically useful structure: tabletops become planes, cans become cylinders. Primitives support grasp synthesis, collision geometry, and scene abstraction where full mesh reconstruction is unnecessary. Why it matters for physical AI: Abstracting clutter into simple shapes gives planners tractable collision models and grasp candidates from partial views, remaining a fast, dependable complement to learned perception in structured picking systems. ### Gimbal Category: Hardware URL: https://almond.bot/glossary/gimbal A gimbal is a pivoted mechanism that permits rotation of a mounted payload about one or more axes, most commonly a motorized two- or three-axis stabilizer that isolates a camera or sensor from the motion of its carrying vehicle using IMU feedback and brushless direct-drive motors. Gimbals stabilize aerial and ground robot cameras, point antennas and sensors, and appear in classical attitude instruments, where the alignment singularity known as gimbal lock originates. Why it matters for physical AI: Stable, actively pointed sensing decouples perception quality from platform motion, improving video for teleoperation and visual policies on drones and rough-terrain robots, while gimbal lock remains the canonical caution about rotation parameterizations. ### Global Planner Category: Navigation & SLAM URL: https://almond.bot/glossary/global-planner A global planner is the navigation component that computes a complete path from a robot's current position to a goal using a map of the environment, employing algorithms such as A*, Dijkstra, or sampling-based planners over costmaps or roadmaps. It operates at coarse timescales and hands its path to a local planner, which handles dynamic obstacles and kinodynamic feasibility in the robot's immediate surroundings. The split is standard in stacks like ROS Navigation and Nav2. Why it matters for physical AI: Layered planning keeps long-horizon route reasoning tractable while reacting quickly to the unexpected; the architecture persists even as learned navigation policies begin to absorb the local layer. ### Goal Image Conditioning Category: Robot Learning URL: https://almond.bot/glossary/goal-image-conditioning Goal image conditioning is a policy conditioning scheme in which the desired task outcome is specified by an image of the goal state, such as a photograph of the assembled object or the target scene arrangement, and the policy learns to act so the observation comes to match it. Goal images sidestep language ambiguity and reward engineering, can be generated by video-prediction or subgoal models as in SuSIE, and enable hindsight relabeling of any trajectory. Why it matters for physical AI: Goal images provide dense, self-supervised task specification that scales with unlabeled data, though capturing a goal photo at deployment is often awkward, so many systems combine image goals with language. ### Goal-Conditioned Policy Category: Robot Learning URL: https://almond.bot/glossary/goal-conditioned-policy A goal-conditioned policy is a control policy that takes a goal specification, such as a target state, image, point, or language command, as an additional input, so a single network can perform many tasks by varying the goal. Training approaches include goal-conditioned behavior cloning, reinforcement learning with universal value functions, and hindsight relabeling, where achieved outcomes are retroactively treated as intended goals to densify supervision. Why it matters for physical AI: Generalist robots must execute arbitrary commands rather than one hardwired task; goal conditioning is the basic mechanism that turns a policy into an instructable system and underlies every language-directed robot model. ### Goods-to-Person (G2P) Category: Applications URL: https://almond.bot/glossary/goods-to-person Goods-to-person is a warehouse fulfillment paradigm in which automated systems deliver inventory to stationary human pickers, rather than workers walking aisles to retrieve items. Implementations include mobile robots carrying shelving pods, popularized by Kiva Systems (acquired by Amazon in 2012 to become Amazon Robotics), shuttle-based automated storage and retrieval systems, and cube storage like AutoStore. G2P multiplies picker productivity and shrinks facility walking distances. Why it matters for physical AI: G2P stations concentrate the remaining human labor into repetitive picking at a fixed workcell, exactly the structured, high-volume setting where learned robotic piece-picking is being deployed to close the automation loop. ### GPU Computing Category: Hardware URL: https://almond.bot/glossary/gpu-computing GPU computing is the use of graphics processing units as massively parallel processors for general computation, programmed through frameworks such as CUDA. In robotics, GPUs train and serve neural policies, accelerate perception, and run massively parallel physics simulation: Isaac Gym demonstrated training locomotion policies with thousands of simulated robots on a single GPU. Embedded GPU modules like the NVIDIA Jetson family bring inference onboard power-constrained robots. Why it matters for physical AI: The entire modern robot learning pipeline, from parallel simulation through foundation-model training to onboard policy inference, is built on GPU throughput; available FLOPs largely set the pace of physical AI progress. ### GR00T (NVIDIA Isaac GR00T) Category: Foundation Models URL: https://almond.bot/glossary/gr00t GR00T is NVIDIA's project and model family for generalist humanoid robot foundation models, announced in 2024 and realized in open models beginning with GR00T N1 in 2025, a vision-language-action model with a dual-system design pairing a vision-language module with a diffusion-based action head for continuous control. The ecosystem includes GR00T-Dreams and GR00T-Mimic synthetic data generation workflows and tight integration with Isaac Sim and the Jetson Thor compute platform. Why it matters for physical AI: Openly released VLA weights plus a synthetic-data-and-simulation toolchain lower the entry cost for humanoid learning research, and platform plays like GR00T signal how compute vendors are positioning around physical AI. ### Gradient Clipping Category: Robot Learning URL: https://almond.bot/glossary/gradient-clipping Gradient clipping is a training technique that limits the magnitude of gradients before a parameter update, most commonly by rescaling the global gradient norm to a threshold, preventing rare large gradients from destabilizing optimization. Originally motivated by exploding gradients in recurrent networks, it is now near-universal in training large transformers and reinforcement learning systems, where nonstationary data and value bootstrapping produce heavy-tailed gradient noise. Why it matters for physical AI: Stable optimization is a precondition for the long, expensive training runs behind robot foundation models, and clipping is one of the small set of defaults that keep large-scale policy training from diverging. ### Graph Neural Network (GNN) Category: Robot Learning URL: https://almond.bot/glossary/graph-neural-network A graph neural network is a neural architecture that operates on graph-structured data by iteratively passing and aggregating messages along edges, producing node, edge, or graph-level representations that respect the graph's relational structure. Robotics applications include learned physics simulators such as DeepMind's graph network simulators, policies that generalize across robot morphologies by treating the kinematic tree as a graph, multi-robot coordination, and reasoning over object-centric scene graphs. Why it matters for physical AI: Robots, scenes, and contact dynamics are naturally relational; graph-structured models offer a route to policies and simulators that generalize across morphologies and object counts where fixed-size networks cannot. ### Graph-Based SLAM Category: Navigation & SLAM URL: https://almond.bot/glossary/graph-based-slam Graph-based SLAM is a formulation of simultaneous localization and mapping as inference over a graph whose nodes are robot poses (and optionally landmarks) and whose edges encode relative measurements from odometry, scan matching, or loop closures. The map and trajectory are recovered by nonlinear least-squares optimization over the graph, as in g2o, Ceres, or GTSAM backends. It superseded filter-based SLAM for most applications due to accuracy and scalability. Why it matters for physical AI: Consistent long-term maps require correcting drift when places are revisited, and graph optimization is the machinery that makes city-scale, lifelong mapping computationally feasible on deployed robots. ### Grasp Affordance Category: Manipulation URL: https://almond.bot/glossary/grasp-affordance A grasp affordance is the property of an object or object region that indicates how it can be grasped, grounding Gibson's ecological notion of affordance in manipulation: handles afford wrapping, rims afford pinching, flat faces afford suction. Computationally, affordance models predict dense per-point or per-pixel graspability and often task-conditioned regions, learned from demonstration, simulation, or large visual datasets, and increasingly from vision-language models that link affordances to semantics. Why it matters for physical AI: Task-aware grasping requires knowing not just where a grasp is stable but which grasp serves the goal, such as holding a knife by the handle to cut; affordance prediction bridges perception and purposeful manipulation. ### Grasp Detection Category: Manipulation URL: https://almond.bot/glossary/grasp-detection Grasp detection is the perception task of predicting viable grasp configurations directly from sensor data such as images, depth maps, or point clouds, without requiring a prior object model. Landmark systems include Dex-Net, which learned grasp robustness from millions of simulated grasps, GG-CNN for real-time pixel-wise grasp maps, and GraspNet and AnyGrasp for dense 6-DoF grasp proposals in clutter. Detected grasps are typically ranked by predicted quality and filtered for reachability. Why it matters for physical AI: Model-free grasping of never-before-seen objects is the entry ticket to unstructured environments like homes and warehouses, and grasp detectors remain core components even inside pipelines led by end-to-end policies. ### Grasp Planning Category: Manipulation URL: https://almond.bot/glossary/grasp-planning Grasp planning is the problem of selecting where and how a robot hand should grasp an object, producing contact points or a gripper pose together with an approach motion, subject to stability criteria, hand kinematics, reachability, collisions, and task requirements. Classical analytic planners optimized force-closure metrics over object models, as in the GraspIt! simulator, while modern systems dominantly use learned grasp detection combined with motion planning for approach and retrieval. Why it matters for physical AI: A grasp is the precondition for nearly every manipulation task, and planning failures cascade downstream; the field's move from analytic models to learned prediction previews the broader arc of physical AI. ### Grasp Pose Category: Manipulation URL: https://almond.bot/glossary/grasp-pose A grasp pose is the position and orientation of a robot end effector, usually expressed as an SE(3) transform relative to the object or scene, at which closing the gripper is expected to yield a successful grasp, often annotated with a gripper opening width and approach direction. Representations range from planar 4-DoF grasps for top-down picking to full 6-DoF poses required for cluttered or shelf scenarios. Why it matters for physical AI: The grasp pose is the standard interface between perception and control in picking systems, and its parameterization, planar versus full 6-DoF, bounds which objects and scene geometries a system can handle. ### Grasp Quality Metric Category: Manipulation URL: https://almond.bot/glossary/grasp-quality-metric A grasp quality metric is a scalar function that scores how good a candidate grasp is, classically from analytic mechanics, such as the Ferrari-Canny epsilon metric measuring the largest disturbance wrench resistible with bounded contact forces, or volume-based measures of the grasp wrench space. Modern systems increasingly use learned quality estimates, trained on simulated or empirical grasp outcomes as in Dex-Net, which better reflect uncertainty in perception and actuation. Why it matters for physical AI: Ranking candidate grasps well is the difference between a picking cell that works at 99 percent and one that jams hourly; the analytic-versus-learned metric evolution mirrors the field's larger empirical turn. ### GraspNet Category: Data & Benchmarks URL: https://almond.bot/glossary/graspnet GraspNet, specifically GraspNet-1Billion from Fang et al. (2020), is a large-scale benchmark for general object grasping comprising about 97,000 RGB-D images of cluttered scenes with over one billion densely annotated 6-DoF grasp poses, plus a standardized evaluation protocol scoring predicted grasps by analytic quality. The accompanying baseline networks and successors such as AnyGrasp made it a reference point for clutter grasp detection research. Why it matters for physical AI: Dense, standardized grasp benchmarks let the community measure real progress in clutter picking rather than compare demos, and pretrained detectors from such datasets ship inside many practical picking stacks. ### Gravity Compensation Category: Control URL: https://almond.bot/glossary/gravity-compensation Gravity compensation is the application of joint torques that exactly cancel the gravitational loads on a robot's links, computed from the dynamics model as the configuration-dependent gravity vector, so the mechanism behaves as if weightless. It is the minimal model-based term in nearly every manipulator controller, enabling low-gain compliant behavior, kinesthetic teaching where a human freely repositions the arm, and safe zero-torque-like modes; passive designs use springs or counterweights. Why it matters for physical AI: Gravity-compensated arms can be physically guided by hand, which is how much kinesthetic demonstration data is collected, and the compliance it permits is foundational to safe operation around people. ### Gravity Grasp Category: Manipulation URL: https://almond.bot/glossary/gravity-grasp A gravity grasp is a hold in which gravity, rather than actively applied clamping force, keeps the object secured against the hand or supporting surfaces, as when an object rests in an upward-facing palm, hangs from a hook-like finger configuration, or sits cradled between fingers. Such grasps require little actuation force and tolerate imprecise contact, but constrain the feasible orientations and accelerations of subsequent motion. Why it matters for physical AI: Exploiting environmental physics instead of fighting it expands what simple grippers can carry, and reasoning about support-based holds features in bimanual transport and the resourceful, human-like strategies learned policies are beginning to exhibit. ### Gripper Category: Hardware URL: https://almond.bot/glossary/gripper A gripper is the end effector a robot uses to grasp and hold objects, most commonly a two-finger parallel-jaw device, with alternatives including vacuum suction cups, three-finger centric grippers, soft pneumatic fingers, and multi-fingered anthropomorphic hands. Selection trades payload, object diversity, cycle speed, sensing, and cost; parallel jaws plus suction cover most industrial picking, while dexterous hands target in-hand manipulation. Standard research examples include the Robotiq 2F-85. Why it matters for physical AI: The gripper defines a robot's action space at the point of contact, and most large manipulation datasets assume parallel jaws, an embodiment prior that shapes what today's foundation policies can and cannot transfer to. ### Gripper Calibration Category: Hardware URL: https://almond.bot/glossary/gripper-calibration Gripper calibration is the process of identifying the geometric and functional parameters of an end effector so commands and sensing are accurate: the tool center point transform relating gripper frame to robot flange, finger opening width as a function of actuator position, grip force versus commanded effort, and zero offsets of integrated sensors. Combined with hand-eye calibration of cameras, it ensures perceived grasp poses are executed where intended. Why it matters for physical AI: Millimeter-level pose errors from stale calibration silently destroy grasp success rates, and learned policies trained on one calibration state can degrade after hardware swaps, making calibration hygiene an operational necessity for fleets. ### Ground Plane Category: Perception URL: https://almond.bot/glossary/ground-plane The ground plane is the planar model of the floor or terrain surface in a robot's environment, estimated from point clouds, stereo, or IMU-aided vision, commonly via RANSAC plane fitting or by assuming a known camera height. Ground plane estimates support obstacle segmentation by removing floor points, drivable-area detection, scale recovery in monocular vision, and placing detected objects at metrically correct heights. Why it matters for physical AI: Separating ground from obstacle is the first perceptual act of any mobile robot, and errors here propagate directly into navigation safety, making robust ground estimation a quietly critical component. ### Ground Reaction Force (GRF) Category: Locomotion URL: https://almond.bot/glossary/ground-reaction-force Ground reaction force is the force exerted by the ground on a body in contact with it, equal and opposite to the force the body applies to the ground. In legged robotics, GRFs at each foot are the only external forces available to accelerate the robot's center of mass, making their planning and control the essence of locomotion; they are constrained to unilateral contact with friction cones and are measured by foot sensors or estimated from joint torques. Why it matters for physical AI: Every locomotion controller, from ZMP walking to RL policies, is ultimately shaping ground reaction forces; accurate GRF estimation and control determine stability on slippery, soft, or uneven real-world terrain. ### Grounded Language Understanding Category: Foundation Models URL: https://almond.bot/glossary/grounded-language-understanding Grounded language understanding is the capability of connecting linguistic expressions to their referents in the physical world, so that phrases like "the mug left of the laptop" resolve to specific percepts, spatial relations, and feasible actions rather than remaining abstract tokens. In robotics it spans referring-expression resolution, spatial and affordance grounding, and instruction following, pursued historically with structured semantic parsers and now dominantly with vision-language models fine-tuned on embodied data. Why it matters for physical AI: Natural language is the interface through which non-experts will command robots, and grounding failures, confidently manipulating the wrong object, are among the most consequential error modes of language-driven systems. ### GRU (Gated Recurrent Unit) Category: Robot Learning URL: https://almond.bot/glossary/gru A GRU is a recurrent neural network cell, introduced by Cho et al. (2014), that maintains a hidden state updated through learned update and reset gates, capturing temporal dependencies with fewer parameters than the LSTM's separate cell state and three gates. In robotics, GRUs provide compact memory for partially observable control, appearing widely in recurrent reinforcement learning policies for locomotion, where they help infer unobserved quantities like terrain friction or payload from interaction history. Why it matters for physical AI: Real robots never observe full state, and lightweight recurrent memory remains a practical choice for onboard policies that must run at high control rates, even as transformers dominate larger offline-trained models. ### GTSAM Category: Software & Middleware URL: https://almond.bot/glossary/gtsam GTSAM (Georgia Tech Smoothing and Mapping) is an open-source C++ library, with Python bindings, for factor-graph-based estimation, developed by Frank Dellaert's group at Georgia Tech. It provides nonlinear least-squares optimizers, incremental smoothing via iSAM2, and rich factor libraries including IMU preintegration, making it a standard backend for SLAM, visual-inertial odometry, and calibration systems in both research and industry. Why it matters for physical AI: Robust open estimation infrastructure lets robotics teams build reliable localization without reinventing solvers, and factor-graph tooling like GTSAM underpins the state estimates that learned policies and planners silently depend on. ### Gyroscope Category: Hardware URL: https://almond.bot/glossary/gyroscope A gyroscope is a sensor that measures angular velocity about one or more axes; in robotics it is nearly always a MEMS vibrating-structure device packaged with accelerometers in an inertial measurement unit. Gyroscope readings integrate to orientation estimates that drift due to bias instability and noise, so they are fused with accelerometers, magnetometers, or vision in attitude filters and visual-inertial odometry. Higher grades, such as fiber-optic gyros, serve demanding navigation applications. Why it matters for physical AI: High-rate rotational sensing is indispensable for balance control in legged robots, camera stabilization, and inertial odometry; gyro noise characteristics directly bound how long a robot can navigate without external position fixes. ## Glossary: H ### H-Infinity Control Category: Control URL: https://almond.bot/glossary/h-infinity-control H-infinity control is a robust control synthesis method that designs controllers by minimizing the H-infinity norm of the closed-loop transfer function from disturbances to regulated outputs, guaranteeing worst-case performance bounds. Introduced by George Zames in 1981, it frames controller design as an optimization problem that explicitly accounts for model uncertainty and disturbance rejection, unlike LQG methods that assume known Gaussian noise. Why it matters for physical AI: Robots deployed outside labs face payload variation, contact disturbances, and unmodeled dynamics; robust synthesis techniques like H-infinity provide stability guarantees that purely learned controllers typically lack, and inform hybrid learning-plus-control architectures. ### Habitat Category: Simulation URL: https://almond.bot/glossary/habitat Habitat is an open-source simulation platform from Meta AI for training embodied agents in photorealistic reconstructions of indoor environments, capable of rendering thousands of frames per second. Habitat 1.0 (2019) targeted visual navigation using scanned datasets such as Matterport3D and HM3D; Habitat 2.0 added interactive rearrangement tasks, and Habitat 3.0 introduced simulated humanoid avatars for human-robot collaboration research. Why it matters for physical AI: Fast photorealistic simulation lets navigation and rearrangement policies train on years of experience in hours, and Habitat's benchmarks have become standard yardsticks for embodied AI before costly real-robot deployment. ### Hallucination Category: Foundation Models URL: https://almond.bot/glossary/hallucination Hallucination is the tendency of large generative models to produce fluent but factually incorrect or ungrounded outputs, such as referencing objects that are not present in a scene. In robotics, hallucination appears when language-model planners propose infeasible actions or when vision-language-action models misreport scene contents, producing confident commands that do not correspond to physical reality. Why it matters for physical AI: A hallucinated plan step can command a robot to grasp a nonexistent object or violate safety constraints, so grounding model outputs in perception and feasibility checks is a core requirement for deploying foundation models on hardware. ### Hand Tracking Category: Teleoperation URL: https://almond.bot/glossary/hand-tracking Hand tracking is the estimation of human hand pose, including wrist position and finger joint angles, from cameras, gloves, or VR headset sensors. Vision-based systems such as MediaPipe Hands and the inside-out tracking on Meta Quest headsets recover 21-keypoint hand skeletons in real time, which teleoperation systems retarget to robot grippers or dexterous hands. Why it matters for physical AI: Glove-free hand tracking makes teleoperation rigs cheap and portable, enabling large-scale collection of dexterous demonstration data that fuels imitation learning for multi-fingered manipulation. ### Hand-Eye Calibration Category: Perception URL: https://almond.bot/glossary/hand-eye-calibration Hand-eye calibration is the procedure of estimating the rigid transformation between a camera and a robot's kinematic frame, either for a camera mounted on the end-effector (eye-in-hand) or fixed in the workspace (eye-to-hand). It is classically formulated as solving the matrix equation AX = XB from paired robot and camera motions, with the Tsai-Lenz method being a standard solution. Why it matters for physical AI: Accurate hand-eye transforms are what let a robot convert pixels into reachable 3D targets; millimeter-level calibration errors translate directly into failed grasps and insertions in real-world deployment. ### Handover Category: Manipulation URL: https://almond.bot/glossary/handover Handover is the transfer of an object between a human and a robot, spanning robot-to-human and human-to-robot directions. A successful handover requires predicting the partner's reach, selecting a grasp that leaves room for the receiver, timing the approach, and modulating grip force release based on sensed load transfer. It is a long-standing benchmark in physical human-robot interaction research. Why it matters for physical AI: Robots working alongside people in homes, hospitals, and factories constantly exchange tools and objects, making fluent handovers a gating capability for collaborative deployment. ### Haptic Feedback Category: Teleoperation URL: https://almond.bot/glossary/haptic-feedback Haptic feedback is the rendering of force, vibration, or tactile cues to a human operator so they can feel interactions occurring at a remote or virtual robot. In bilateral teleoperation, forces measured or estimated at the follower robot are reflected to the leader device, closing a kinesthetic loop that improves performance on contact-rich tasks like insertion and suturing. Why it matters for physical AI: Operators with force feedback produce gentler, more successful contact-rich demonstrations, improving both teleoperated task completion and the quality of data collected for training manipulation policies. ### Hardware-in-the-Loop (HIL) Category: Simulation URL: https://almond.bot/glossary/hardware-in-the-loop Hardware-in-the-loop testing is a validation methodology in which real hardware components, typically embedded controllers or actuators, run against a simulated model of the rest of the system in real time. The physical device receives synthetic sensor signals and its outputs drive the simulation, exposing timing, interface, and firmware issues that pure software simulation misses. Why it matters for physical AI: HIL rigs catch integration bugs in motor drivers, communication buses, and safety logic before they damage expensive robots, and they enable regression testing of control stacks without tying up full systems. ### Harmonic Drive Category: Hardware URL: https://almond.bot/glossary/harmonic-drive A harmonic drive is a strain-wave gear consisting of an elliptical wave generator, a flexible toothed flexspline, and a rigid circular spline, achieving high reduction ratios (typically 50:1 to 160:1) in a compact, coaxial package with near-zero backlash. Its light weight and precision made it the dominant gearing in collaborative robot and humanoid joints, though it exhibits torsional compliance, friction, and limited backdrivability. Why it matters for physical AI: Gearbox choice shapes what policies can do: harmonic drives give precise position control but poor torque transparency, pushing designs toward joint torque sensors or alternative transmissions for contact-rich learned skills. ### HDF5 (Hierarchical Data Format v5) Category: Data & Benchmarks URL: https://almond.bot/glossary/hdf5 HDF5 is a binary file format and library for storing large, heterogeneous numerical data in a hierarchical structure of groups and datasets, with support for chunking, compression, and partial I/O. Robot learning frameworks such as robomimic and many teleoperation pipelines store demonstrations as HDF5 files containing synchronized image streams, proprioception, and actions. Why it matters for physical AI: Training manipulation policies means streaming terabytes of multimodal demonstration data; efficient chunked formats like HDF5 determine whether dataloaders keep GPUs fed during large-scale policy training. ### Heat Pipe Category: Hardware URL: https://almond.bot/glossary/heat-pipe A heat pipe is a sealed, passive two-phase heat transfer device in which a working fluid evaporates at the hot end, condenses at the cool end, and returns through a capillary wick, moving heat with effective conductivity far exceeding solid copper. Robots use heat pipes and vapor chambers to cool onboard GPUs and motor drivers inside sealed, fanless enclosures. Why it matters for physical AI: Onboard inference for large policies pushes embedded GPUs toward their thermal limits inside compact robot bodies, and passive heat-pipe cooling sustains compute performance without dust-prone fans in field environments. ### Hello Robot Stretch Category: Hardware URL: https://almond.bot/glossary/hello-robot-stretch Hello Robot Stretch is a lightweight mobile manipulator combining a differential-drive base, a vertical lift, and a horizontally telescoping arm with a compliant wrist and gripper. Designed by Hello Robot for research and assistive applications, its slim footprint, low mass, and comparatively low cost made it a popular platform for home manipulation research and benchmarks such as open-vocabulary mobile manipulation. Why it matters for physical AI: Affordable, safe mobile manipulators lower the barrier to collecting household manipulation data and testing policies in real homes, a setting most industrial arms cannot enter. ### Hertz (Control Rate) Category: Control URL: https://almond.bot/glossary/hertz Hertz is the SI unit of frequency, one cycle per second, used in robotics to specify the rates of nested control and perception loops. Typical stacks run joint-level servo loops at 1 kHz or faster, whole-body or Cartesian controllers at 100 Hz to 1 kHz, cameras at 30 to 60 Hz, and learned policy inference anywhere from a few hertz to around 50 Hz. Why it matters for physical AI: Rate mismatches between slow policy inference and fast actuator loops are bridged by action chunking and interpolation; insufficient control frequency causes instability and poor contact behavior in deployed systems. ### Heteroscedastic Uncertainty Category: Robot Learning URL: https://almond.bot/glossary/heteroscedastic-uncertainty Heteroscedastic uncertainty is input-dependent noise, where the variance of a predicted quantity changes across the input space rather than remaining constant (homoscedastic). Models capture it by predicting a variance alongside each mean output and training with a negative log-likelihood loss, letting the network report higher aleatoric uncertainty for ambiguous observations such as occluded objects. Why it matters for physical AI: Policies that know when their predictions are unreliable can slow down, gather information, or request help, which is central to safe autonomy and to deciding when human intervention is needed. ### Hexapod Category: Locomotion URL: https://almond.bot/glossary/hexapod A hexapod is a six-legged walking robot, a morphology that permits statically stable gaits such as the alternating tripod, in which three legs always form a support polygon. Inspired by insect locomotion, hexapods handle rough, unstructured terrain without dynamic balance control. The term also denotes six-strut Stewart platform positioners used for precision motion, a distinct mechanism. Why it matters for physical AI: Static stability makes hexapods forgiving testbeds for terrain-adaptive gait learning, and their redundancy tolerates leg failures, which is valuable for inspection missions in hazardous environments. ### Hierarchical Policy Category: Robot Learning URL: https://almond.bot/glossary/hierarchical-policy A hierarchical policy is a control architecture in which a high-level policy selects subgoals, skills, or latent commands at a coarse timescale, and one or more low-level policies translate them into motor actions at a fine timescale. Modern instantiations pair a slow vision-language model for reasoning with a fast reactive controller, as in dual-system architectures like Figure's Helix and NVIDIA GR00T N1. Why it matters for physical AI: Splitting deliberation from reflexes lets large models reason at a few hertz while motor control runs at hundreds, making long-horizon tasks tractable and inference budgets manageable on real robots. ### Hierarchical Reinforcement Learning (HRL) Category: Robot Learning URL: https://almond.bot/glossary/hierarchical-reinforcement-learning Hierarchical reinforcement learning is a family of RL methods that decompose long-horizon problems using temporal abstraction, learning high-level policies that invoke temporally extended lower-level behaviors. The options framework of Sutton, Precup, and Singh formalized skills as options with initiation sets and termination conditions; later approaches include feudal networks and goal-conditioned hierarchies such as HIRO. Why it matters for physical AI: Sparse rewards over hundreds of steps cripple flat RL; temporal abstraction shortens the effective horizon, improving exploration and credit assignment for multi-stage manipulation and locomotion tasks. ### HIL-SERL Category: Robot Learning URL: https://almond.bot/glossary/hil-serl HIL-SERL (Human-in-the-Loop Sample-Efficient Reinforcement Learning) is a framework from UC Berkeley (Luo et al., 2024) for training manipulation policies with real-world RL, combining off-policy actor-critic learning seeded with demonstrations and corrective human interventions during training. It reported near-perfect success rates on contact-rich tasks such as insertion and assembly within roughly one to two and a half hours of robot time. Why it matters for physical AI: Demonstrating that on-robot RL can reach production-grade reliability in hours rather than weeks makes autonomous fine-tuning a practical stage after imitation pretraining for deployed manipulation systems. ### Hindsight Experience Replay (HER) Category: Robot Learning URL: https://almond.bot/glossary/hindsight-experience-replay Hindsight Experience Replay is a technique for goal-conditioned reinforcement learning, introduced by Andrychowicz et al. (2017), that relabels failed trajectories as successes for the goals the agent actually reached. By storing each transition with substituted achieved goals, HER extracts dense learning signal from sparse binary rewards, and it enabled off-policy algorithms like DDPG to solve robotic pushing, sliding, and pick-and-place benchmarks. Why it matters for physical AI: Reward engineering is a major bottleneck for robot RL; hindsight relabeling turns every attempt into useful data, improving sample efficiency where real-robot experience is expensive and rewards are naturally sparse. ### Hollow Shaft Category: Hardware URL: https://almond.bot/glossary/hollow-shaft A hollow shaft is a rotating shaft with a central bore, used in robot joints and actuators so that cables, pneumatic lines, or optical fibers can be routed through the axis of rotation rather than around it. Hollow-shaft harmonic drives, frameless motors, and slip rings are common in serial arms, keeping wiring internal and eliminating external cable loops that snag or fatigue. Why it matters for physical AI: Internal cable routing is what allows continuous-rotation joints and clean arm exteriors, directly improving reliability and workspace safety in robots that run around the clock in cluttered environments. ### Home Robot Category: Applications URL: https://almond.bot/glossary/home-robot A home robot is a robot designed to operate autonomously in domestic environments, ranging from single-task appliances such as robotic vacuums to general-purpose mobile manipulators intended for tidying, laundry, and kitchen work. Homes are a canonical unstructured setting, with high scene diversity, deformable objects, pets, and people, and benchmarks such as Meta's HomeRobot open-vocabulary mobile manipulation challenge target this domain. Why it matters for physical AI: The home is the ultimate generalization test: no two are alike, so household robots are a driving application for foundation-model policies trained on diverse cross-embodiment data. ### Homogeneous Transformation Category: Math & Kinematics URL: https://almond.bot/glossary/homogeneous-transformation A homogeneous transformation is a 4x4 matrix representing a rigid-body pose in SE(3), combining a 3x3 rotation matrix and a 3x1 translation vector, with a bottom row of [0 0 0 1]. Transformations compose by matrix multiplication and invert analytically, which makes them the standard machinery for chaining coordinate frames through a robot's kinematic tree and for expressing camera and object poses. Why it matters for physical AI: Every pose in a robot software stack, from forward kinematics to calibration to grasp targets, is a homogeneous transform, and frame bookkeeping errors remain one of the most common sources of deployment bugs. ### Homography Category: Perception URL: https://almond.bot/glossary/homography A homography is a 3x3 projective transformation relating pixel coordinates of a plane viewed from two camera poses, or of any scene viewed by a purely rotating camera. It can be estimated from four or more point correspondences, typically with RANSAC for robustness, and underlies image stitching, planar visual servoing, and ground-plane mapping in mobile robotics. Why it matters for physical AI: Many workspaces are dominated by planes, tabletops, floors, walls, so homography-based alignment offers a cheap, robust geometric prior for visual servoing and camera-to-workspace registration without full 3D reconstruction. ### Horizon (Task Horizon) Category: Robot Learning URL: https://almond.bot/glossary/horizon Horizon is the number of decision steps or the duration over which a task or episode unfolds, and by extension the temporal span a policy or planner must reason over. Long-horizon tasks, such as multi-stage assembly or meal preparation, compound per-step errors multiplicatively and dilute credit assignment, which is why they remain substantially harder than short-horizon skills like single grasps. Why it matters for physical AI: Compounding error growth with horizon length is the central obstacle to reliable multi-stage autonomy, motivating hierarchical policies, subgoal decomposition, and recovery behaviors in deployed systems. ### Hospital Robot Category: Applications URL: https://almond.bot/glossary/hospital-robot A hospital robot is a robot deployed in clinical environments, spanning logistics platforms that ferry medications, linens, and lab samples (such as Aethon TUG and Diligent's Moxi), UV disinfection robots, telepresence units, and surgical systems like the da Vinci. These robots must navigate crowded corridors, operate elevators, and meet stringent hygiene and safety requirements around patients and staff. Why it matters for physical AI: Healthcare combines chronic labor shortages with highly structured logistics, making hospitals an early proving ground for mobile robots operating safely among untrained members of the public. ### Human Pose Estimation Category: Perception URL: https://almond.bot/glossary/human-pose-estimation Human pose estimation is the task of recovering the configuration of the human body from sensor data, as 2D or 3D skeletal keypoints or as a parametric mesh such as SMPL. Systems like OpenPose established real-time multi-person 2D estimation, while modern methods regress full 3D body shape from monocular video, enabling motion capture without markers. Why it matters for physical AI: Extracting body motion from ordinary video unlocks human demonstrations at internet scale for humanoid learning, and real-time pose tracking underpins collision avoidance and intent prediction around people. ### Human-in-the-Loop Category: Robot Learning URL: https://almond.bot/glossary/human-in-the-loop Human-in-the-loop refers to system designs in which a person actively participates in a robot's learning or operation, providing demonstrations, corrective interventions, preference labels, or approvals rather than leaving the system fully autonomous. In robot learning it spans interactive imitation methods like DAgger and HG-DAgger, intervention-based RL such as HIL-SERL, and deployment-time oversight where operators take over on failure. Why it matters for physical AI: Interventions both guarantee safe behavior during the fragile early life of a policy and generate precisely targeted training data, forming the improvement flywheel behind most commercial robot fleets. ### Human-Robot Interaction (HRI) Category: Applications URL: https://almond.bot/glossary/human-robot-interaction Human-robot interaction is the multidisciplinary field studying how robots and people communicate, collaborate, and share space, drawing on robotics, cognitive science, design, and psychology. It covers physical HRI, where safety during contact is paramount, and social dimensions such as legibility of robot motion, trust calibration, and natural-language communication, with dedicated venues like the ACM/IEEE HRI conference. Why it matters for physical AI: General-purpose robots will work among people who never read a manual, so legible motion, intent communication, and calibrated trust are as decisive for adoption as raw task success rates. ### Humanoid Foundation Model Category: Foundation Models URL: https://almond.bot/glossary/humanoid-foundation-model A humanoid foundation model is a large pretrained policy intended to provide general-purpose control for humanoid robots, mapping vision, language, and proprioception to whole-body or upper-body actions across many tasks. Examples include NVIDIA's GR00T N1, trained on a pyramid of internet video, synthetic, and teleoperated robot data, and Figure's Helix, which pairs a vision-language backbone with a high-rate visuomotor controller. Why it matters for physical AI: A single model that transfers skills across tasks and humanoid platforms would amortize data collection across an entire fleet, which is the core economic bet behind current humanoid programs. ### Humanoid Robot Category: Hardware URL: https://almond.bot/glossary/humanoid-robot A humanoid robot is a robot with a human-like body plan, typically a torso, two arms, two legs (or a wheeled base), and a head, sized to operate in spaces built for people. Contemporary platforms include Boston Dynamics' electric Atlas, Tesla Optimus, Figure's robots, Agility's Digit, and Unitree's G1 and H1, most combining learned locomotion controllers with manipulation policies. Why it matters for physical AI: Matching the human form factor lets one platform reuse buildings, tools, and workflows designed for people, and it maximizes transfer from human motion data to robot control. ### Hybrid Force-Position Control Category: Control URL: https://almond.bot/glossary/hybrid-force-position-control Hybrid force-position control is a control scheme, formalized by Raibert and Craig in 1981, that partitions the task space into orthogonal directions, controlling position along unconstrained axes and force along axes constrained by contact. Selection matrices assign each Cartesian direction to one mode, so a robot can, for example, regulate normal force against a surface while tracking a trajectory tangential to it. Why it matters for physical AI: Tasks like polishing, wiping, and insertion inherently mix motion and force objectives; hybrid control provides the classical template that learned contact-rich policies and compliant controllers still build upon. ### Hybrid Grasp Category: Manipulation URL: https://almond.bot/glossary/hybrid-grasp A hybrid grasp is a grasp that combines characteristics of power and precision grasping, for example using the palm and proximal fingers to secure an object while leaving distal fingertips free to control or reorient it. Grasp taxonomies such as Cutkosky's and the GRASP taxonomy catalogue such intermediate postures, which humans use extensively when handling tools. Why it matters for physical AI: Tool use frequently demands both a secure hold and fine fingertip control at once, so multi-fingered hands and their policies must master hybrid postures to move beyond simple pick-and-place. ### Hybrid Imitation RL Category: Robot Learning URL: https://almond.bot/glossary/hybrid-imitation-rl Hybrid imitation RL is the family of methods that combine demonstrations with reinforcement learning, either by pretraining a policy with behavior cloning and fine-tuning it with RL, by placing demonstrations in the replay buffer of an off-policy learner (as in DDPGfD, DAPG, and RLPD), or by shaping rewards from demonstration state distributions. The combination addresses RL's exploration problem while allowing the policy to exceed demonstrator performance. Why it matters for physical AI: Imitation gets policies working quickly but plateaus at human quality and coverage; layering RL on top pushes success rates toward the near-perfect reliability commercial deployments require. ### Hydraulic Actuator Category: Hardware URL: https://almond.bot/glossary/hydraulic-actuator A hydraulic actuator is an actuator that converts pressurized fluid flow into linear or rotary motion, offering exceptional force and power density along with inherent shock tolerance. Hydraulics powered the original Boston Dynamics Atlas and BigDog as well as most heavy construction machinery, but leaks, plumbing complexity, noise, and poor energy efficiency have driven most new humanoid and legged designs to electric actuation. Why it matters for physical AI: The industry-wide shift from hydraulic to electric actuation traded peak force density for cleanliness, efficiency, and the precise torque control that learned controllers exploit, reshaping legged robot design. ### Hysteresis Category: Hardware URL: https://almond.bot/glossary/hysteresis Hysteresis is the dependence of a system's output on its input history, producing different responses on loading versus unloading paths. In robots it appears as backlash and torsional wind-up in gearboxes, friction and stretch in tendon drives, and lag in sensors and elastomer-based tactile skins, introducing path-dependent errors that simple linear models cannot capture. Why it matters for physical AI: Unmodeled hysteresis silently corrupts calibration, repeatability, and sim-to-real transfer, so identifying or learning these effects is often the difference between a policy working in simulation and on hardware. ## Glossary: I ### I2C Protocol Category: Hardware URL: https://almond.bot/glossary/i2c-protocol I2C (Inter-Integrated Circuit) is a synchronous, two-wire serial bus using a shared data line (SDA) and clock line (SCL), where a controller addresses multiple peripheral devices over the same pair of wires. Standard speeds range from 100 kHz to 3.4 MHz, and it is the default interface for short-range sensor wiring such as IMUs, magnetic encoders, and time-of-flight rangefinders on robot electronics. Why it matters for physical AI: Low-cost robot builds and sensor-dense end-effectors lean heavily on I2C for wiring simplicity, though its limited speed and cable length push high-rate joints toward CAN or EtherCAT. ### Image Encoder Category: Foundation Models URL: https://almond.bot/glossary/image-encoder An image encoder is a neural network that maps raw images to compact feature embeddings consumed by downstream policy or planning heads, with ResNet and Vision Transformer (ViT) architectures being the most common backbones. Robot policies frequently reuse encoders pretrained at internet scale, such as CLIP, SigLIP, or DINOv2, either frozen or fine-tuned, rather than learning vision from robot data alone. Why it matters for physical AI: Pretrained visual representations inject broad semantic knowledge that small robot datasets cannot provide, substantially improving generalization of manipulation policies to novel objects and scenes. ### ImageNet Category: Data & Benchmarks URL: https://almond.bot/glossary/imagenet ImageNet is a large-scale image dataset organized by the WordNet hierarchy, whose 1000-class ILSVRC classification benchmark catalyzed the deep learning era after AlexNet's 2012 result. Beyond the benchmark itself, ImageNet pretraining became the default initialization for perception backbones, and "an ImageNet moment for robotics" is shorthand for the field's search for similarly transformative shared datasets, motivating efforts like Open X-Embodiment. Why it matters for physical AI: The dataset's history frames the central question of robot learning: whether pooled, diverse robot data can do for manipulation what ImageNet did for computer vision. ### Imitation Learning Category: Robot Learning URL: https://almond.bot/glossary/imitation-learning Imitation learning is the paradigm of training policies from expert demonstrations rather than from reward signals, most simply via behavior cloning, which regresses actions from observations, and more robustly via interactive methods like DAgger or distribution-matching approaches like GAIL. Teleoperated demonstration collection combined with expressive policy classes, as in ALOHA's ACT and Diffusion Policy, drives most current real-robot manipulation results. Why it matters for physical AI: Demonstrations sidestep reward engineering and unsafe exploration, making imitation the workhorse for bootstrapping manipulation skills and the pretraining substrate for robot foundation models. ### Impedance Control Category: Control URL: https://almond.bot/glossary/impedance-control Impedance control is a control strategy, introduced by Neville Hogan in 1985, that regulates the dynamic relationship between a robot's motion and interaction forces, making the end-effector or joints behave like a programmable mass-spring-damper system. Rather than tracking positions or forces exactly, the robot yields predictably to contact, with stiffness and damping parameters tuned per task or per direction. Why it matters for physical AI: Compliant behavior under contact is fundamental to safe operation around humans and to contact-rich skills, and learned policies increasingly output impedance targets rather than raw positions for robustness. ### Impedance Matching Category: Hardware URL: https://almond.bot/glossary/impedance-matching Impedance matching is the practice of matching a source's impedance to its load to maximize power transfer or minimize signal reflection. In robot electronics it governs termination of high-speed buses and RF antennas; the mechanical analogue is selecting a transmission ratio so a motor's inertia and torque characteristics match the load, maximizing acceleration or energy efficiency at the joint. Why it matters for physical AI: Gear ratio selection as mechanical impedance matching determines whether legs can absorb impacts and arms can be backdriven, shaping the actuation trade-offs behind dynamic, contact-tolerant robots. ### Implicit Q-Learning (IQL) Category: Robot Learning URL: https://almond.bot/glossary/implicit-q-learning Implicit Q-Learning is an offline reinforcement learning algorithm (Kostrikov, Nair, and Levine, 2021) that avoids querying the value of out-of-distribution actions by fitting a state-value function with expectile regression, then extracting a policy via advantage-weighted regression. Its stability and simplicity made it a standard baseline for offline RL benchmarks like D4RL and a common choice for learning from static robot datasets. Why it matters for physical AI: Robots accumulate large logs of mixed-quality experience that cannot be safely re-explored; offline algorithms like IQL turn those static datasets into improved policies without on-robot trial and error. ### In-Context Learning Category: Foundation Models URL: https://almond.bot/glossary/in-context-learning In-context learning is the ability of large pretrained models to adapt to new tasks from examples or instructions placed in their input prompt, without any weight updates. In robotics, the paradigm appears as policies conditioned on a handful of demonstration trajectories at inference time, as in one-shot imitation learning, and as LLM planners steered by few-shot examples of task decompositions. Why it matters for physical AI: Adapting a deployed robot by showing it a demonstration, rather than launching a fine-tuning run, would collapse the cost of teaching new tasks from days to minutes. ### In-Hand Manipulation Category: Manipulation URL: https://almond.bot/glossary/in-hand-manipulation In-hand manipulation is the repositioning or reorientation of an object within a hand without setting it down or regrasping with the other hand, using finger gaiting, rolling, and controlled sliding. OpenAI's Dactyl demonstrated learned in-hand cube reorientation with a Shadow Hand trained in simulation with domain randomization, and subsequent work has achieved similar skills with cheaper hands and touch sensing. Why it matters for physical AI: Tool use routinely requires adjusting grip mid-task, so in-hand skills are a prerequisite for general manipulation and a stress test for sim-to-real transfer of dexterous policies. ### Incremental Learning Category: Robot Learning URL: https://almond.bot/glossary/incremental-learning Incremental learning is the continual updating of a model as new data arrives, without retraining from scratch and without catastrophically forgetting previously acquired capabilities. Techniques include regularization toward old weights, replay of stored or generated past data, and modular architectures that add capacity per task, all balancing plasticity for new skills against stability of old ones. Why it matters for physical AI: Deployed fleets encounter new objects, layouts, and tasks continuously; robots that can fold this experience into their policies without regressing existing skills compound in value over time. ### Industrial Robot Category: Applications URL: https://almond.bot/glossary/industrial-robot An industrial robot is, per ISO 8373, an automatically controlled, reprogrammable, multipurpose manipulator programmable in three or more axes, used in applications such as welding, painting, palletizing, machine tending, and assembly. Millions of units operate worldwide, dominated by six-axis arms from vendors like FANUC, ABB, KUKA, and Yaskawa, traditionally running fixed programs inside safety-fenced cells. Why it matters for physical AI: Industrial arms represent an enormous installed base of reliable hardware whose economics change dramatically if foundation-model policies can replace fixed programming with rapid, general task adaptation. ### Inertia Tensor Category: Math & Kinematics URL: https://almond.bot/glossary/inertia-tensor The inertia tensor is a symmetric 3x3 matrix describing how a rigid body's mass is distributed about a reference point, determining the torque required for angular acceleration about any axis. Its eigenvectors define principal axes, and accurate link inertia values, specified in robot description formats like URDF, are required for dynamics computations, feedforward control, and physically plausible simulation. Why it matters for physical AI: Wrong inertia parameters corrupt simulated dynamics and torque-level controllers alike, making inertial identification a quiet but critical ingredient of sim-to-real transfer and dynamic whole-body control. ### Inertial Measurement Unit (IMU) Category: Hardware URL: https://almond.bot/glossary/inertial-measurement-unit An inertial measurement unit is a sensor package combining a three-axis accelerometer and three-axis gyroscope, often with a magnetometer, measuring specific force and angular velocity at rates from hundreds to thousands of hertz. Because integrating its measurements drifts over time, IMU data is fused with cameras, LiDAR, or leg odometry via Kalman or complementary filters to produce stable orientation and velocity estimates. Why it matters for physical AI: High-rate inertial sensing is the backbone of state estimation for legged balance, drone stabilization, and visual-inertial odometry, bridging the gaps between slower exteroceptive sensor updates. ### Inference Latency Category: Foundation Models URL: https://almond.bot/glossary/inference-latency Inference latency is the time between a model receiving an observation and producing its output, which for robot policies bounds the achievable control rate and reaction time. Billion-parameter vision-language-action models can take tens to hundreds of milliseconds per forward pass, so systems mitigate latency with action chunking, asynchronous execution, quantization, distillation, and hierarchical designs that reserve large models for slow deliberation. Why it matters for physical AI: A policy that reasons brilliantly but reacts slowly fails at dynamic tasks, so the latency-capability trade-off dictates model size, hardware selection, and architecture for every deployed learned controller. ### Information Matrix Category: Navigation & SLAM URL: https://almond.bot/glossary/information-matrix The information matrix is the inverse of a covariance matrix, representing the certainty of a Gaussian estimate; large entries indicate strong constraints between the corresponding variables. Graph-based SLAM exploits the natural sparsity of the information form, since each measurement links only a few poses and landmarks, enabling efficient factor-graph solvers such as g2o and GTSAM to optimize millions of variables. Why it matters for physical AI: Sparse information-form representations are why modern SLAM scales to long deployments, letting robots maintain consistent maps and pose estimates over hours of operation on modest onboard compute. ### Insertion Task Category: Manipulation URL: https://almond.bot/glossary/insertion-task An insertion task is the class of manipulation problems requiring one part to be fitted into another under tight clearances, with peg-in-hole as the canonical example and connector mating, gear meshing, and furniture assembly as practical variants. Success under sub-millimeter tolerances demands exploiting contact through compliance, force feedback, or learned search strategies, and benchmarks such as the NIST Assembly Task Boards standardize evaluation. Why it matters for physical AI: Insertion dominates electronics assembly and manufacturing work, and because vision alone rarely achieves the required precision, it is the proving ground for force-aware and contact-rich learned policies. ### Inspection Robot Category: Applications URL: https://almond.bot/glossary/inspection-robot An inspection robot is a robot that autonomously surveys infrastructure or industrial assets, capturing imagery, thermal readings, gas concentrations, or ultrasonic thickness measurements from pipelines, power lines, tanks, and plant floors. Quadrupeds such as Boston Dynamics Spot and ANYbotics ANYmal patrol facilities on scheduled missions, while specialized crawlers and drones access pipes, vessels, and heights unsafe for humans. Why it matters for physical AI: Inspection offers repetitive, hazardous, well-scoped missions with clear ROI, making it one of the first commercially successful applications of autonomous legged and aerial robots. ### Instance Segmentation Category: Perception URL: https://almond.bot/glossary/instance-segmentation Instance segmentation is the computer vision task of detecting each object in an image and delineating it with a pixel-accurate mask, distinguishing individual instances of the same class. Mask R-CNN established the dominant two-stage approach, and the Segment Anything Model (SAM) introduced promptable, class-agnostic segmentation that transfers zero-shot to novel objects and domains. Why it matters for physical AI: Per-object masks let manipulation systems isolate a grasp target from clutter, estimate object extents, and track items through a scene, forming the perceptual front end of most pick-and-place stacks. ### Instruction Following Category: Foundation Models URL: https://almond.bot/glossary/instruction-following Instruction following is the capability of a robot to execute tasks specified in natural language, requiring the system to ground words in perception and action. Language-conditioned policies such as RT-1 and RT-2 map instructions directly to motor commands, planners like SayCan decompose instructions into affordance-weighted skills, and benchmarks such as CALVIN evaluate long-horizon language-conditioned manipulation. Why it matters for physical AI: Natural language is the interface that lets non-experts retask robots on the fly, and instruction-following generalization is the headline capability distinguishing robot foundation models from single-task policies. ### Intel RealSense Category: Hardware URL: https://almond.bot/glossary/intel-realsense Intel RealSense is a family of compact depth cameras, most prominently the D400 series, which computes depth from active-IR-assisted stereo and streams aligned RGB-D imagery over USB at up to 90 frames per second. Models like the D435 and D455 became near-default sensors on research manipulation and mobile robot platforms due to their low cost, small size, and mature SDK. Why it matters for physical AI: Ubiquitous, affordable RGB-D sensing standardized the observation space of a generation of manipulation datasets and policies, and camera choice still shapes cross-lab data compatibility today. ### Intent Recognition Category: Perception URL: https://almond.bot/glossary/intent-recognition Intent recognition is the inference of a human's goal or next action from observable cues such as body motion, gaze, hand trajectories, and speech. Approaches range from probabilistic models over goal hypotheses to learned sequence models that predict reach targets or activity labels, giving robots the lead time needed to act proactively rather than reactively. Why it matters for physical AI: Anticipating what a person will do next enables fluent collaboration, timely handovers, and preemptive collision avoidance, capabilities that separate genuinely collaborative robots from merely co-located ones. ### Interaction Data Category: Data & Benchmarks URL: https://almond.bot/glossary/interaction-data Interaction data is data generated by an agent acting in an environment, pairing observations with the actions taken and their outcomes, as opposed to passive data such as internet video that lacks action labels. Robot interaction data from teleoperation, autonomous rollouts, and interventions carries the causal, closed-loop structure policies need, but it is orders of magnitude scarcer than passive web data. Why it matters for physical AI: The scarcity of action-labeled experience is the defining bottleneck of robot learning, driving investment in teleoperation fleets, simulation-generated rollouts, and deployment flywheels that convert operations into training data. ### Interactive Imitation Learning Category: Robot Learning URL: https://almond.bot/glossary/interactive-imitation-learning Interactive imitation learning is a family of methods in which an expert provides feedback during policy execution, rather than only supplying demonstrations upfront. DAgger established the paradigm by querying expert actions on states the learner visits, and variants like HG-DAgger and intervention-based schemes let humans take over only when the policy errs, concentrating supervision where it is most informative. Why it matters for physical AI: Covariate shift makes pure behavior cloning brittle on states of its own making; on-policy corrections repair exactly those failure modes, and intervention logs double as a natural deployment supervision signal. ### Interactive Perception Category: Perception URL: https://almond.bot/glossary/interactive-perception Interactive perception is the paradigm in which a robot acts on the world specifically to improve its perception, such as pushing objects apart to segment clutter, lifting an item to gauge its weight, or moving the camera to resolve an occlusion. By exploiting the coupling between action and sensing, it recovers properties like mass, articulation, and rigidity that passive observation cannot reveal. Why it matters for physical AI: Embodiment is an informational advantage: robots that probe and nudge the world resolve ambiguities that stall vision-only systems, which is essential in cluttered, partially observable real environments. ### Interpolation Category: Math & Kinematics URL: https://almond.bot/glossary/interpolation Interpolation is the computation of intermediate values between known points, used throughout robotics to turn sparse waypoints or low-rate commands into smooth, dense trajectories. Common schemes include linear and cubic spline interpolation for joint positions, minimum-jerk profiles for smoothness, and spherical linear interpolation (slerp) for rotations, which respects the geometry of orientation space. Why it matters for physical AI: Learned policies typically emit actions far below servo rates, so interpolation quality directly affects motion smoothness, tracking fidelity, and hardware wear when executing action chunks on real robots. ### Intrinsic Motivation Category: Robot Learning URL: https://almond.bot/glossary/intrinsic-motivation Intrinsic motivation is the use of internally generated reward signals, such as curiosity, novelty, prediction error, or empowerment, to drive exploration in the absence of external task rewards. Methods like the Intrinsic Curiosity Module (ICM) and Random Network Distillation (RND) reward visiting hard-to-predict states, enabling agents to acquire diverse behaviors and data in sparse-reward or reward-free settings. Why it matters for physical AI: Autonomous data collection at fleet scale cannot rely on hand-specified rewards for every skill, so intrinsic objectives offer a path to robots that practice and expand their own competence. ### Inventory Management Robot Category: Applications URL: https://almond.bot/glossary/inventory-management-robot An inventory management robot is a mobile robot that automates stock counting and shelf auditing in warehouses and retail stores, scanning shelves with cameras, barcode readers, or RFID to detect out-of-stocks, misplaced items, and price discrepancies. Systems range from aisle-roaming units like Simbe's Tally to drones that fly warehouse racking, feeding inventory databases far more frequently than manual counts. Why it matters for physical AI: Shelf scanning is a structured, repetitive perception task where robots already outperform manual audits, and the resulting high-frequency inventory data compounds into better logistics automation downstream. ### Inverse Dynamics Category: Control URL: https://almond.bot/glossary/inverse-dynamics Inverse dynamics is the computation of the joint torques required to produce a specified motion, given joint positions, velocities, and accelerations along with the robot's mass and inertia parameters. The Recursive Newton-Euler Algorithm computes it efficiently for kinematic chains, and it underpins computed-torque control, feedforward compensation, and gravity compensation in torque-controlled robots. Why it matters for physical AI: Feedforward torques from inverse dynamics let controllers track aggressive motions with low feedback gains, enabling the compliant yet precise behavior that dynamic legged and manipulation platforms depend on. ### Inverse Kinematics (IK) Category: Math & Kinematics URL: https://almond.bot/glossary/inverse-kinematics Inverse kinematics is the problem of finding joint configurations that place a robot's end-effector at a desired pose, the inverse of forward kinematics. Solutions may be analytic for specific geometries such as 6-DOF arms with spherical wrists, or numerical, using Jacobian-based iteration or nonlinear optimization as in solvers like TRAC-IK; redundant arms admit infinitely many solutions, and singularities and joint limits complicate solving. Why it matters for physical AI: IK is the translation layer between task-space intent, whether from planners, teleoperators, or learned policies, and the joint commands hardware executes, running inside nearly every manipulation stack. ### Inverse Reinforcement Learning (IRL) Category: Robot Learning URL: https://almond.bot/glossary/inverse-reinforcement-learning Inverse reinforcement learning is the problem of inferring the reward function an expert is optimizing from observations of its behavior, formalized by Ng and Russell in 2000. Maximum-entropy IRL (Ziebart et al., 2008) resolved the ambiguity among consistent rewards probabilistically, and adversarial descendants such as GAIL match the expert's state-action distribution without explicitly recovering a reward. Why it matters for physical AI: Recovering the objective behind demonstrations yields policies that generalize beyond copied trajectories and transfers human preferences into reward signals usable for further reinforcement learning. ### Inverted Pendulum Category: Control URL: https://almond.bot/glossary/inverted-pendulum An inverted pendulum is a pendulum balanced above its pivot, an inherently unstable system that serves as the canonical benchmark for stabilizing control, from cart-pole in RL curricula to classical state-space design. The Linear Inverted Pendulum Model (LIPM) abstracts a walking robot's center of mass dynamics and underlies ZMP-based gait generation and capture-point analysis for bipeds. Why it matters for physical AI: Reduced-order pendulum models remain the analytical core of humanoid balance and push recovery, complementing learned locomotion policies with interpretable stability criteria for deployment. ### IP Rating (Ingress Protection) Category: Hardware URL: https://almond.bot/glossary/ip-rating An IP rating is a two-digit code defined by IEC 60529 classifying an enclosure's protection against solid particles (first digit, 0-6) and liquids (second digit, 0-9), so IP54 denotes dust-protected and splash-resistant while IP67 denotes dust-tight and immersion-resistant. Robots intended for outdoor, washdown, or dusty industrial environments specify IP ratings for joints, connectors, and electronics enclosures. Why it matters for physical AI: Field deployment fails on mundane physics before algorithms: sealing against dust, rain, and washdown chemicals determines whether a robot survives farms, construction sites, and food plants. ### Isaac Gym Category: Simulation URL: https://almond.bot/glossary/isaac-gym Isaac Gym is NVIDIA's GPU-accelerated physics simulation environment for reinforcement learning that runs thousands of environment instances in parallel on a single GPU, keeping both simulation and policy training on-device to eliminate CPU-GPU transfer bottlenecks. It enabled results such as training ANYmal locomotion policies in minutes (Rudin et al., 2021) and has since been superseded by Isaac Lab. Why it matters for physical AI: Massively parallel simulation collapsed RL wall-clock time from weeks to minutes, making iterative sim-to-real development of locomotion and manipulation policies practical for ordinary research teams. ### Isaac Lab Category: Simulation URL: https://almond.bot/glossary/isaac-lab Isaac Lab is NVIDIA's open-source robot learning framework built on Isaac Sim, unifying the successors of Isaac Gym and the Orbit project. It provides GPU-parallelized environments, tiled multi-camera rendering, actuator and sensor models, and workflows for reinforcement learning, imitation learning, and synthetic data generation across manipulators, quadrupeds, and humanoids. Why it matters for physical AI: A maintained, standardized training framework atop a photorealistic simulator shortens the path from task definition to deployable policy and anchors much of the current humanoid learning ecosystem. ### Isaac ROS Category: Software & Middleware URL: https://almond.bot/glossary/isaac-ros Isaac ROS is NVIDIA's collection of GPU-accelerated ROS 2 packages providing hardware-accelerated perception, including visual SLAM, the nvblox 3D reconstruction library, AprilTag detection, stereo depth, and DNN inference nodes. Its NITROS transport adapts ROS 2 message passing for zero-copy GPU pipelines, targeting Jetson-class onboard computers. Why it matters for physical AI: Perception at frame rate on embedded power budgets is a chronic deployment constraint, and GPU-accelerated middleware lets full autonomy stacks run onboard instead of offloading to servers. ### Isaac Sim Category: Simulation URL: https://almond.bot/glossary/isaac-sim Isaac Sim is NVIDIA's robotics simulator built on the Omniverse platform, combining PhysX physics, RTX ray-traced rendering, and the USD scene description format. It supports ROS integration, digital-twin construction, and synthetic data generation through its Replicator toolkit, and serves as the foundation for Isaac Lab's robot learning workflows. Why it matters for physical AI: Photorealistic rendering plus accurate physics narrows both the visual and dynamics sim-to-real gaps, while synthetic data generation supplies the labeled diversity real-world collection cannot economically match. ### ISO 10218 Category: Safety & Standards URL: https://almond.bot/glossary/iso-10218 ISO 10218 is the international safety standard for industrial robots, with Part 1 covering the robot itself and Part 2 covering robot systems and integration. The 2025 revision substantially reorganized the standard and absorbed the collaborative operation requirements formerly detailed in ISO/TS 15066, specifying requirements for functional safety, protective devices, and collaborative applications. Why it matters for physical AI: Certification against ISO 10218 is the gateway to selling and integrating robots in industrial settings, and its collaborative provisions define the compliance envelope for robots sharing space with workers. ### ISO/TS 15066 Category: Safety & Standards URL: https://almond.bot/glossary/iso-ts-15066 ISO/TS 15066 is a technical specification detailing safety requirements for collaborative robot operation, supplementing ISO 10218. It is best known for its biomechanical limits tables, which specify maximum permissible forces and pressures for quasi-static and transient contact across regions of the human body, operationalizing the power- and force-limiting collaborative mode; its content has been folded into the 2025 revision of ISO 10218. Why it matters for physical AI: The contact force limits give engineers concrete, testable numbers for designing robots that may touch people, shaping actuator sizing, padding, and speed supervision in every cobot deployment. ### Iterative Closest Point (ICP) Category: Perception URL: https://almond.bot/glossary/iterative-closest-point Iterative Closest Point is an algorithm for registering two point clouds by alternating between matching each point to its nearest neighbor in the other cloud and computing the rigid transformation that minimizes the matched distances, iterating to convergence. Introduced by Besl and McKay in 1992, it spawned variants including point-to-plane ICP and Generalized ICP, and it requires a reasonable initial alignment to avoid local minima. Why it matters for physical AI: ICP-style registration powers LiDAR scan matching for odometry, model-based object pose refinement for grasping, and map alignment, making it one of the most-run geometric algorithms on deployed robots. ### Iterative Learning Control Category: Control URL: https://almond.bot/glossary/iterative-learning-control Iterative learning control is a control technique for systems that repeat the same task, updating the feedforward command after each trial based on the error recorded in previous trials. Over successive repetitions the tracking error converges toward zero, compensating for repeatable disturbances and model errors without requiring an accurate plant model. Why it matters for physical AI: Repetitive industrial operations such as dispensing, machining, and pick cycles can approach near-perfect tracking through trial-to-trial refinement, a classical complement to learning-based control that improves with every execution. ## Glossary: J ### Jaco Arm Category: Hardware URL: https://almond.bot/glossary/jaco-arm The Jaco arm is a lightweight, low-power robotic arm developed by Kinova, originally designed as an assistive device mounted on powered wheelchairs to help users with upper-body mobility impairments perform daily tasks. Built with carbon-fiber links and available in six- and seven-DOF configurations with two- or three-finger grippers, it also became a widely used research platform and appears in several public manipulation datasets. Why it matters for physical AI: Assistive arms like Jaco demonstrate manipulation delivering direct human benefit under strict safety and power constraints, and their research adoption seeded demonstration data now folded into cross-embodiment corpora. ### Jacobian Matrix Category: Math & Kinematics URL: https://almond.bot/glossary/jacobian-matrix The Jacobian matrix in robotics is the matrix of partial derivatives mapping joint velocities to end-effector linear and angular velocity, with each column giving the end-effector twist produced by unit motion of one joint. Its transpose maps end-effector wrenches to joint torques, its rank drops at kinematic singularities, and its pseudoinverse drives differential inverse kinematics and operational-space control. Why it matters for physical AI: Jacobian computations run inside virtually every manipulator controller, converting task-space commands from planners or learned policies into joint motion while exposing singularities and manipulability limits. ### Jerk Category: Math & Kinematics URL: https://almond.bot/glossary/jerk Jerk is the third time derivative of position, the rate of change of acceleration, measured in meters per second cubed or the angular equivalent. High jerk excites structural vibration, degrades tracking accuracy, stresses transmissions, and produces motion that humans perceive as abrupt, so trajectory optimization frequently penalizes or bounds jerk, and minimum-jerk profiles characterize smooth human reaching movements. Why it matters for physical AI: Smoothness is not cosmetic: jerk-aware trajectories reduce wear, protect payloads and sensors, and produce the predictable motion people working near robots implicitly rely on. ### Jerk Limit Category: Control URL: https://almond.bot/glossary/jerk-limit A jerk limit is a bound on the third derivative of position enforced during trajectory generation, producing S-curve velocity profiles instead of trapezoidal ones with instantaneous acceleration steps. Online trajectory generators such as Ruckig compute time-optimal jerk-limited motions in real time, allowing controllers to react to new targets while respecting kinematic constraints of the hardware. Why it matters for physical AI: Enforcing jerk limits at the trajectory layer lets aggressive learned or reactive commands execute without exciting resonances or tripping drives, bridging policy outputs and hardware tolerance in deployed systems. ### Jetson (NVIDIA) Category: Hardware URL: https://almond.bot/glossary/jetson Jetson is NVIDIA's line of embedded computing modules pairing ARM CPUs with CUDA-capable GPUs and dedicated deep learning accelerators, spanning power envelopes from the entry-level Orin Nano to the high-end AGX Orin and the humanoid-oriented Jetson Thor. Jetson modules are the dominant onboard compute for robots running neural perception and policy inference without a tether to external servers. Why it matters for physical AI: Onboard GPU compute determines how large a policy can run at what control rate, making the Jetson roadmap a de facto constraint on deployable model architectures for mobile robots. ### Joint Coupling Category: Hardware URL: https://almond.bot/glossary/joint-coupling Joint coupling is mechanical or kinematic interdependence between joints, where actuating one degree of freedom produces motion or torque at another. It arises by design in differential wrists, tendon-driven fingers, and underactuated hands, where fewer motors drive more joints, and as a parasitic effect in cable-routed transmissions; either way it requires a coupling matrix or model in the kinematics and control stack. Why it matters for physical AI: Coupled, underactuated designs cut cost and weight in hands and wrists, but their nonlinear motor-to-joint mappings complicate calibration, simulation fidelity, and the action spaces exposed to learned policies. ### Joint Limits Category: Math & Kinematics URL: https://almond.bot/glossary/joint-limits Joint limits are the bounds on each joint's allowable range, primarily position limits set by mechanical stops or wiring, along with velocity, acceleration, and torque bounds specified by the manufacturer. They are encoded in robot description files, enforced in software ahead of hard stops, and act as constraints in inverse kinematics, motion planning, and trajectory optimization. Why it matters for physical AI: Policies and planners that ignore joint limits produce unreachable or self-damaging commands, so limit-aware action spaces and safety layers are baseline requirements for running learned controllers on hardware. ### Joint Position Control Category: Control URL: https://almond.bot/glossary/joint-position-control Joint position control is the servo-level tracking of commanded joint positions, conventionally implemented as cascaded position, velocity, and current loops with PID controllers running at kilohertz rates on the motor drives. It is the stiffest and most widespread control interface on commercial arms, and many learned manipulation policies output joint position targets that these loops execute. Why it matters for physical AI: Position control's stiffness gives repeatability but punishes unexpected contact, so the choice between position, velocity, and torque interfaces shapes both policy action spaces and contact safety. ### Joint Space Category: Math & Kinematics URL: https://almond.bot/glossary/joint-space Joint space is the coordinate space whose axes are a robot's joint variables, so a configuration of an n-joint arm is a point in an n-dimensional space, in contrast to task or Cartesian space, which describes the end-effector pose. Motion planning, collision checking, and most learned policies operate in joint space, with forward kinematics mapping it to task space. Why it matters for physical AI: Whether policies act in joint space or task space is a recurring design decision, trading direct hardware correspondence and singularity-free execution against intuitive, embodiment-transferable commands. ### Joint Space Trajectory Category: Control URL: https://almond.bot/glossary/joint-space-trajectory A joint space trajectory is a time-parameterized path specified directly in joint coordinates, typically as spline segments through waypoints with velocity, acceleration, and jerk constraints. Executing in joint space avoids online inverse kinematics and is immune to Cartesian singularities, though the resulting end-effector path between waypoints is not a straight line and must be checked for collisions. Why it matters for physical AI: Most industrial motion and much learned-policy execution ultimately compiles to joint space trajectories, whose smoothness and timing determine cycle time, tracking accuracy, and mechanical wear. ### Joint Torque Category: Control URL: https://almond.bot/glossary/joint-torque Joint torque is the rotational effort exerted at a robot joint, produced by the actuator and transmission and balanced against gravity, inertia, friction, and external contact forces. It can be estimated from motor current or measured directly with strain-gauge sensors or series elastic elements, and torque-level control interfaces enable impedance control, gravity compensation, and sensitive collision detection. Why it matters for physical AI: Access to accurate joint torque is what separates robots that can feel and comply from those that only track positions, underpinning contact-rich skills and safe physical interaction with people. ### Junction Box Category: Hardware URL: https://almond.bot/glossary/junction-box A junction box is a protective enclosure where electrical cables are joined, distributed, and terminated, providing strain relief, environmental sealing, and organized access for maintenance. In robot cells and mobile platforms, junction boxes consolidate power distribution, emergency-stop wiring, and signal routing between the robot, its controller, and peripheral devices, often built to IP-rated and electrical-code requirements. Why it matters for physical AI: Fleet reliability lives and dies on connectors and wiring; disciplined junction and distribution design cuts intermittent faults, simplifies field service, and keeps safety circuits verifiable at scale. ## Glossary: K ### Kalman Filter Category: Navigation & SLAM URL: https://almond.bot/glossary/kalman-filter The Kalman filter is a recursive state estimator, introduced by Rudolf Kalman in 1960, that optimally fuses a linear dynamics prediction with noisy measurements under Gaussian assumptions, alternating predict and update steps. The extended (EKF) and unscented (UKF) variants handle nonlinear systems by linearization or sigma-point propagation, and Kalman-family filters remain the standard tool for fusing IMU, odometry, GPS, and vision. Why it matters for physical AI: Nearly every robot's belief about where it is and how fast it is moving flows through a Kalman-style filter, making estimation quality a silent ceiling on control and navigation performance. ### Keyframe Selection Category: Robot Learning URL: https://almond.bot/glossary/keyframe-selection Keyframe selection is the identification of a sparse set of salient frames or waypoints that summarize a trajectory, demonstration, or video. In robot learning, keyframe-based imitation trains policies to predict bottleneck end-effector poses, as in systems like C2FARM and PerAct, shortening effective horizons; in visual SLAM, keyframes are the subset of camera frames retained for mapping and bundle adjustment. Why it matters for physical AI: Predicting a handful of decisive waypoints instead of dense action streams improves sample efficiency and long-horizon reliability, a recurring trick behind data-efficient manipulation learners. ### Keypoint Detection Category: Perception URL: https://almond.bot/glossary/keypoint-detection Keypoint detection is the localization of distinctive points in images, spanning classical corner and blob detectors such as SIFT and ORB used for matching and SLAM, learned detectors like SuperPoint, and semantic keypoints denoting functional object parts. Keypoint-based manipulation methods such as kPAM represent objects by task-relevant points, providing a compact, category-level state that transfers across object instances. Why it matters for physical AI: Sparse keypoints offer an interpretable middle ground between raw pixels and full pose, giving manipulation policies category-level generalization and giving SLAM its geometric anchors. ### Kidnapped Robot Problem Category: Navigation & SLAM URL: https://almond.bot/glossary/kidnapped-robot-problem The kidnapped robot problem is the localization scenario in which a robot is moved to an unknown location without notification, so its belief about its pose is confidently wrong rather than merely uncertain. It is the standard stress test for global relocalization, favoring methods like Monte Carlo localization with particle reinjection and place-recognition systems that can recover from total tracking failure. Why it matters for physical AI: Real deployments produce kidnapping-like events constantly, through elevator rides, manual repositioning, and tracking dropouts, so graceful global relocalization separates robust fleets from demo-grade navigation. ### Kinect Category: Hardware URL: https://almond.bot/glossary/kinect Kinect is Microsoft's line of consumer RGB-D cameras, launched in 2010 for the Xbox using structured-light depth sensing, followed by the time-of-flight Kinect v2 and the Azure Kinect developer kit. Its low price brought dense depth sensing to robotics labs at scale, catalyzing a wave of research in 3D mapping (including KinectFusion), skeleton tracking, and RGB-D perception. Why it matters for physical AI: Kinect's commoditization of depth sensing reshaped robot perception research and established the RGB-D observation paradigm that today's manipulation datasets and sensors still follow. ### Kinematic Chain Category: Math & Kinematics URL: https://almond.bot/glossary/kinematic-chain A kinematic chain is an assembly of rigid links connected by joints, the structural abstraction underlying manipulators and legged robots. Chains may be serial, as in most robot arms, parallel, as in Delta robots and Stewart platforms, or branched, as in humanoids with multiple limbs from a common torso, and robot description formats like URDF encode them as trees of links and joints. Why it matters for physical AI: The chain abstraction is what makes kinematics, dynamics, and simulation computable for arbitrary morphologies, letting one software stack serve arms, hands, quadrupeds, and humanoids alike. ### Kinematic Redundancy Category: Math & Kinematics URL: https://almond.bot/glossary/kinematic-redundancy Kinematic redundancy is the property of a robot having more degrees of freedom than its task requires, as in a seven-DOF arm performing a six-DOF pose task. The surplus creates a null space of joint motions that leave the end-effector fixed, which controllers exploit for secondary objectives such as obstacle and singularity avoidance, joint-limit centering, or posture optimization. Why it matters for physical AI: Redundant arms can keep elbows clear of humans and obstacles while holding a task pose, providing the workspace flexibility that cluttered, human-shared environments demand of deployed manipulators. ### Kinematics Category: Math & Kinematics URL: https://almond.bot/glossary/kinematics Kinematics is the study of motion, position, velocity, and acceleration, without regard to the forces that cause it. Robot kinematics encompasses forward kinematics, mapping joint values to end-effector pose; inverse kinematics, the reverse problem; and differential kinematics, relating joint and task-space velocities through the Jacobian, with conventions like Denavit-Hartenberg parameters standardizing the modeling. Why it matters for physical AI: Kinematic models are the shared geometric language between planners, controllers, simulators, and learned policies; their accuracy bounds calibration quality and every downstream precision claim. ### Kinesthetic Teaching Category: Robot Learning URL: https://almond.bot/glossary/kinesthetic-teaching Kinesthetic teaching is a demonstration method in which a human physically guides the robot's body through a task while the robot records its own joint trajectories, typically with the arm in a gravity-compensated, freely backdrivable mode. It requires no retargeting since demonstrations are captured directly in the robot's embodiment, though the teacher's presence occludes cameras and hand-guiding multi-limb systems is awkward. Why it matters for physical AI: Hand-guiding is the fastest way for non-programmers to teach single-arm skills, and its embodiment-matched trajectories feed directly into imitation learning and skill libraries without correspondence problems. ### Kinova Category: Hardware URL: https://almond.bot/glossary/kinova Kinova is a Canadian robotics manufacturer known for lightweight, low-power robotic arms, including the assistive Jaco arm and the Gen3 research platform, a seven-DOF arm with integrated joint torque sensing and an optional wrist-mounted vision module. Its arms are widespread in academic manipulation research and appear in multiple public datasets and benchmark setups. Why it matters for physical AI: Accessible, torque-sensing research arms like Kinova's lowered the hardware barrier for manipulation learning research, and their prevalence makes them a recurring embodiment in cross-robot datasets. ### KL Divergence Category: Math & Kinematics URL: https://almond.bot/glossary/kl-divergence KL divergence (Kullback-Leibler divergence) is an asymmetric measure of how one probability distribution differs from a reference distribution, zero only when they match. In robot learning it constrains policy updates in trust-region and proximal methods like TRPO and PPO, regularizes latent spaces in variational autoencoders, keeps fine-tuned policies close to pretrained ones, and defines the distillation loss between teacher and student models. Why it matters for physical AI: Bounding distribution shift with KL terms is how RL fine-tuning avoids destroying pretrained competence, a stabilizing mechanism now central to adapting robot foundation models safely. ### Knot Tying Category: Manipulation URL: https://almond.bot/glossary/knot-tying Knot tying is a benchmark manipulation task requiring a robot to form knots in rope or thread, combining deformable linear object perception, topological state reasoning, and precise bimanual coordination over long horizons. It has served as a testbed from early surgical suturing systems to modern learned approaches that track rope state from vision and plan sequences of grasps and crossings. Why it matters for physical AI: Deformable, self-occluding objects with topology-dependent goals defeat rigid-body assumptions, so progress on knot tying signals maturity in bimanual, contact-rich manipulation of the soft materials filling homes and hospitals. ### Knowledge Distillation Category: Foundation Models URL: https://almond.bot/glossary/knowledge-distillation Knowledge distillation is the training of a smaller student model to reproduce the outputs or intermediate representations of a larger teacher, introduced by Hinton, Vinyals, and Dean in 2015 using softened output distributions. In robotics it compresses large vision-language-action models for edge deployment and appears as teacher-student policy distillation, where a privileged-observation teacher trained in simulation supervises a vision-based student. Why it matters for physical AI: Distillation reconciles the scale that makes foundation policies capable with the latency and power budgets of onboard compute, and privileged-teacher schemes are a standard route to deployable sim-trained controllers. ### Koch v1.1 Arm Category: Hardware URL: https://almond.bot/glossary/koch-v11-arm The Koch v1.1 arm is a low-cost, open-source robotic arm design by Alexander Koch, built from 3D-printed parts and Dynamixel servomotors and typically assembled as a leader-follower pair for teleoperated demonstration collection. Popularized through Hugging Face's LeRobot ecosystem, it became a common entry point for hobbyists and researchers experimenting with imitation learning on hardware costing a few hundred dollars. Why it matters for physical AI: Sub-thousand-dollar teleoperation hardware radically widened who can collect real demonstration data, seeding a community-scale supply of manipulation datasets and reproducible robot learning experiments. ### KUKA iiwa Category: Hardware URL: https://almond.bot/glossary/kuka-iiwa The KUKA LBR iiwa is a seven-DOF collaborative robot arm with joint torque sensors in every axis, offered in 7 kg and 14 kg payload variants. Its integrated torque sensing supports impedance control, hand-guiding, and sensitive collision detection, and it became a fixture of research labs, notably as the platform for the Amazon Picking Challenge entries and Google's early large-scale grasping experiments. Why it matters for physical AI: Torque-sensing arms like the iiwa proved that whole-arm compliance and collision sensitivity could be industrial-grade, setting the hardware template for safe physical interaction that learned contact-rich skills now build on. ## Glossary: L ### L2 Regularization Category: Robot Learning URL: https://almond.bot/glossary/l2-regularization L2 regularization is a technique that penalizes large parameter values by adding the sum of squared weights, scaled by a coefficient, to a model's training loss. It encourages smaller, more evenly distributed weights, which reduces overfitting and improves generalization. In deep learning it is closely related to weight decay, though the two differ subtly under adaptive optimizers such as Adam, motivating the decoupled formulation used in AdamW. Why it matters for physical AI: Robot policies are often trained on small, expensive demonstration datasets, so regularization is a primary defense against memorizing training trajectories instead of learning behaviors that generalize to new scenes and object placements. ### Lagrangian Dynamics Category: Math & Kinematics URL: https://almond.bot/glossary/lagrangian-dynamics Lagrangian dynamics is a formulation of classical mechanics that derives a system's equations of motion from the Lagrangian, defined as kinetic energy minus potential energy, via the Euler-Lagrange equations. For robot manipulators it yields the standard form M(q)q̈ + C(q, q̇)q̇ + g(q) = τ, relating joint torques to motion through the mass matrix, Coriolis terms, and gravity vector. It is often more systematic than Newton-Euler methods for deriving closed-form dynamics. Why it matters for physical AI: Accurate dynamics models underpin torque control, gravity compensation, and physics simulators used to train policies, and they anchor system identification when transferring simulation-trained controllers onto real hardware. ### Language Grounding Category: Foundation Models URL: https://almond.bot/glossary/language-grounding Language grounding is the problem of connecting words and phrases to their referents in the physical world, such as objects, spatial relations, and actions perceivable by a robot. Grounded understanding lets a system map an instruction like "put the mug left of the plate" onto specific percepts and motor commands. Approaches range from classical semantic parsing to vision-language models such as CLIP that align text and image embeddings in a shared space. Why it matters for physical AI: Instruction-following robots only work if language reliably maps to real objects and feasible actions; grounding failures are a dominant error mode for vision-language-action models deployed in cluttered, open-world environments. ### Language-Conditioned Policy Category: Robot Learning URL: https://almond.bot/glossary/language-conditioned-policy A language-conditioned policy is a robot control policy that takes a natural-language instruction as an input alongside observations, producing actions that accomplish the commanded task. Instructions are typically encoded with a pretrained language or vision-language model and fused with visual features, as in BC-Z, RT-1, and RT-2. A single network can thereby perform many tasks, with the instruction selecting the behavior at inference time. Why it matters for physical AI: Language is the most practical task-specification interface for general-purpose robots, letting one foundation model cover thousands of tasks and enabling non-experts to retask hardware without programming or new demonstrations. ### Last Mile Delivery Category: Applications URL: https://almond.bot/glossary/last-mile-delivery Last mile delivery is the final segment of a logistics chain, moving goods from a local hub to the end customer, and is typically the most expensive and labor-intensive leg. Robotic approaches include sidewalk delivery robots such as those from Starship Technologies, delivery drones, and autonomous vehicles. These systems must handle pedestrian-rich, unstructured outdoor environments with GPS-denied stretches, curbs, and weather. Why it matters for physical AI: Sidewalk-scale autonomy stresses perception, localization, and safety around uninstructed humans, making last mile delivery a proving ground for deploying navigation stacks at consumer scale and price points. ### Latency Category: Control URL: https://almond.bot/glossary/latency Latency is the time delay between a cause and its observable effect in a system, such as the interval between a sensor reading and the corresponding actuator response. In robotics it accumulates across sensing, perception, policy inference, communication, and actuation stages. Excessive or variable latency degrades closed-loop stability, limits achievable control bandwidth, and makes contact-rich or dynamic tasks difficult to execute reliably. Why it matters for physical AI: Large learned policies can take tens of milliseconds per inference, so budgeting latency across perception and control determines whether a foundation model can run reactive, contact-rich manipulation on real hardware. ### Latency Compensation Category: Teleoperation URL: https://almond.bot/glossary/latency-compensation Latency compensation is the set of techniques used to mitigate the destabilizing effects of time delay in a control loop, particularly in teleoperation over networks. Methods include predictive displays that render forecasted robot state to the operator, Smith predictors that internally simulate the delayed plant, and passivity-based approaches such as wave variables that preserve stability in delayed bilateral teleoperation. Why it matters for physical AI: Remote teleoperation is both a deployment mode and a data-collection engine for robot learning; compensating delay keeps demonstrations smooth and safe even when operators are far from the hardware. ### Latent Action Representation Category: Robot Learning URL: https://almond.bot/glossary/latent-action-representation A latent action representation is a learned, compact encoding of actions or action-like transitions, often inferred from data without ground-truth control labels. Systems such as LAPA and Genie learn discrete latent actions directly from videos by modeling frame-to-frame changes, enabling pretraining on action-free web video. The latent actions are later decoded or aligned to real robot commands with a small amount of labeled data. Why it matters for physical AI: Human video vastly outnumbers robot demonstration data; latent actions let foundation models mine that video for transferable manipulation knowledge, then bridge to executable controls with modest robot-specific fine-tuning. ### Latent Diffusion Category: Foundation Models URL: https://almond.bot/glossary/latent-diffusion Latent diffusion is a generative modeling approach that runs the diffusion denoising process in the compressed latent space of a pretrained autoencoder rather than in raw pixel space. Introduced by Rombach et al. in 2022 and popularized by Stable Diffusion, it dramatically reduces the compute needed for high-resolution image and video generation while preserving quality. Conditioning mechanisms allow generation guided by text, images, or other modalities. Why it matters for physical AI: Efficient generative models power video prediction, world models, and synthetic data pipelines for robotics, where generating plausible future observations at practical compute budgets is essential for planning and large-scale training. ### Latent Space Category: Robot Learning URL: https://almond.bot/glossary/latent-space A latent space is a learned, typically lower-dimensional vector space in which a model represents data, with coordinates capturing abstract factors of variation rather than raw measurements. Encoders such as VAEs, contrastive models, and autoencoding transformers map images, states, or trajectories into latent vectors that support downstream prediction, planning, or control. Distances and directions in a well-structured latent space often correspond to semantically meaningful changes. Why it matters for physical AI: Planning and control in latent space, as in world-model agents like Dreamer, sidesteps raw-pixel dynamics and lets robots reason compactly about high-dimensional camera observations in real time. ### Layer Normalization Category: Robot Learning URL: https://almond.bot/glossary/layer-normalization Layer normalization is a neural network technique, introduced by Ba et al. in 2016, that normalizes the activations of a layer across the feature dimension for each individual sample, then applies a learned scale and shift. Unlike batch normalization it is independent of batch size and behaves identically at training and inference time, making it well suited to transformers and recurrent networks, where it is now standard. Why it matters for physical AI: Stable normalization is one reason transformer-based robot policies train reliably; it also suits the small or streaming batch regimes common in on-robot learning where batch statistics are unreliable. ### Lazy Grasping Category: Manipulation URL: https://almond.bot/glossary/lazy-grasping Lazy grasping is a strategy that defers committing to a specific grasp pose or fully evaluating grasp candidates until execution demands it, refining the choice online as the robot approaches the object. Analogous to lazy evaluation in motion planning, it avoids exhaustive up-front grasp analysis, instead validating only the candidates actually pursued and re-ranking them as new observations arrive during the reach. Why it matters for physical AI: Deferring grasp commitment lets closed-loop policies exploit fresh, close-range observations, improving robustness to perception error, object movement, and clutter compared with executing a grasp planned once from a distance. ### Lead Screw Category: Hardware URL: https://almond.bot/glossary/lead-screw A lead screw is a mechanical linear actuator component that converts rotary motion into linear motion through the sliding contact of a threaded screw and nut, commonly with trapezoidal or Acme thread profiles. Compared with ball screws, lead screws are cheaper, quieter, and often self-locking under load, but exhibit higher friction, lower efficiency, and more backlash and wear. They are widely used in 3D printers, linear stages, and low-cost robot axes. Why it matters for physical AI: Actuator choices like lead screws versus ball screws set the cost, backdrivability, and precision envelope of affordable robot hardware, directly shaping what manipulation behaviors a platform can learn and execute. ### LEAP Hand Category: Hardware URL: https://almond.bot/glossary/leap-hand The LEAP Hand is a low-cost, open-source anthropomorphic robot hand developed at Carnegie Mellon University and released in 2023, designed for machine learning research in dexterous manipulation. It uses 16 direct-driven hobby servo motors in a four-finger configuration, costs on the order of two thousand dollars to build, and its kinematics allow all four fingertip poses useful for common grasps. Assembly instructions, CAD, and control code are fully open. Why it matters for physical AI: Dexterous manipulation research was long gated by hands costing six figures; accessible open hardware lets many labs collect real-world dexterity data, accelerating learned in-hand manipulation and multi-finger grasping. ### Learned Reward Category: Robot Learning URL: https://almond.bot/glossary/learned-reward A learned reward is a reward function acquired from data rather than hand-engineered, typically inferred from demonstrations via inverse reinforcement learning, from human preference comparisons, or from success classifiers trained on labeled outcomes. Vision-language models are increasingly used to produce reward signals zero-shot by scoring how well observations match a task description. The learned reward then supervises reinforcement learning or trajectory ranking. Why it matters for physical AI: Hand-specifying rewards for contact-rich, real-world tasks is brittle and slow; learned rewards let robots improve autonomously from raw experience, a key ingredient for scaling reinforcement learning beyond simulation. ### Learning from Demonstration (LfD) Category: Robot Learning URL: https://almond.bot/glossary/learning-from-demonstration Learning from demonstration (LfD) is the umbrella paradigm in which robots acquire skills from examples provided by a teacher, most often via teleoperation or kinesthetic guidance, instead of explicit programming or trial-and-error reward optimization. Methods span behavior cloning, dynamic movement primitives, inverse reinforcement learning, and modern transformer or diffusion policies trained on demonstration corpora such as those collected with ALOHA-style rigs. Why it matters for physical AI: Demonstrations remain the dominant supervision source for robot foundation models; nearly every large-scale manipulation dataset, from RT-1 to Open X-Embodiment, is built on human-provided examples of successful task execution. ### Learning from Observation Category: Robot Learning URL: https://almond.bot/glossary/learning-from-observation Learning from observation is imitation learning from state or video sequences that lack action labels, requiring the agent to infer how to reproduce observed behavior without knowing the exact commands the demonstrator issued. Techniques include learning inverse dynamics models to recover pseudo-actions, matching state-occupancy distributions as in GAIfO, and extracting latent actions from raw video. Human videos are the most important application domain. Why it matters for physical AI: Internet-scale human video contains far more manipulation experience than all robot datasets combined; action-free imitation is the main route to tapping it for pretraining generalist robot policies. ### Learning from Play Category: Robot Learning URL: https://almond.bot/glossary/learning-from-play Learning from play is a data collection and training paradigm that uses long, unsegmented streams of goal-free, curiosity-driven teleoperated interaction rather than task-scripted demonstrations. Introduced prominently in Lynch et al.'s 2019 Play-LMP work, play data densely covers the state space and captures diverse ways to manipulate an environment; goal- or language-conditioned policies are then trained by relabeling reached states as intended goals. Why it matters for physical AI: Play data is cheap to collect, naturally diverse, and reusable across many tasks via hindsight relabeling, making it an efficient substrate for training multi-task and goal-conditioned manipulation policies. ### Learning Rate Category: Robot Learning URL: https://almond.bot/glossary/learning-rate The learning rate is the scalar hyperparameter that controls the step size of gradient-based optimization, determining how much model parameters change per update. Values that are too large cause divergence or oscillation, while values too small slow convergence or trap training in poor regions. Practical training combines warmup phases, decay schedules such as cosine annealing, and adaptive optimizers like Adam that rescale steps per parameter. Why it matters for physical AI: Fine-tuning pretrained robot foundation models is acutely learning-rate sensitive: overly aggressive updates destroy pretrained visual representations, while conservative ones fail to adapt the policy to new embodiments and tasks. ### Legged Robot Category: Locomotion URL: https://almond.bot/glossary/legged-robot A legged robot is a mobile robot that locomotes using articulated limbs rather than wheels or tracks, spanning bipeds, quadrupeds such as ANYmal and Boston Dynamics' Spot, and humanoids. Legs enable traversal of stairs, rubble, and discontinuous terrain inaccessible to wheels, at the cost of high-dimensional, contact-rich control. Modern controllers are commonly trained with reinforcement learning in simulation and transferred to hardware. Why it matters for physical AI: Legged platforms proved that sim-to-real reinforcement learning can produce robust real-world behavior at scale, and they extend robot deployment into human-built and outdoor spaces designed for legs, not wheels. ### LeKiwi Category: Hardware URL: https://almond.bot/glossary/lekiwi LeKiwi is a low-cost, open-source mobile manipulator design in the Hugging Face LeRobot ecosystem, combining an SO-100/SO-101 class arm with a three-omniwheel holonomic "kiwi drive" base built from hobby servos and 3D-printed parts. The platform targets a build cost in the hundreds of dollars and integrates directly with LeRobot's teleoperation, dataset recording, and policy training tooling for mobile manipulation research. Why it matters for physical AI: Mobile manipulation data has been scarce because platforms were expensive; commodity open-source bases let a broad community collect and share whole-robot demonstrations, widening the data funnel for generalist policies. ### LeRobot Category: Software & Middleware URL: https://almond.bot/glossary/lerobot LeRobot is an open-source robot learning library from Hugging Face, launched in 2024, that provides end-to-end tooling for real-world imitation learning: teleoperated data recording, a standardized dataset format, and training and inference code for policies including ACT, Diffusion Policy, and ported vision-language-action models. It targets affordable hardware such as the SO-100 and SO-101 arms and hosts community datasets and pretrained weights on the Hugging Face Hub. Why it matters for physical AI: By standardizing data formats and lowering hardware costs, a shared open stack lets thousands of practitioners contribute demonstrations and reproduce results, mirroring the open-source dynamics that accelerated language models. ### LeRobot Dataset Format Category: Data & Benchmarks URL: https://almond.bot/glossary/lerobot-dataset-format The LeRobot dataset format is the standardized storage schema used by the LeRobot library for robot demonstration data, organizing synchronized camera streams, proprioceptive states, and actions into episodes with per-frame timestamps and metadata. Tabular data is stored in Parquet with video frames encoded as compressed MP4 for efficient loading, and datasets are versioned and shared through the Hugging Face Hub for direct streaming into training pipelines. Why it matters for physical AI: Fragmented, incompatible logging formats have long impeded data pooling across labs; a common schema makes community datasets interoperable, which is a precondition for training cross-embodiment robot foundation models. ### LIBERO Category: Data & Benchmarks URL: https://almond.bot/glossary/libero LIBERO is a simulation benchmark for lifelong robot learning in manipulation, introduced in 2023, comprising 130 language-conditioned tabletop tasks built on the Robosuite/MuJoCo stack. Its task suites, including LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-100, isolate different axes of knowledge transfer such as spatial relations, object categories, and task goals. It ships with human demonstration datasets and is widely used to evaluate vision-language-action models and continual learning methods. Why it matters for physical AI: Standardized multi-task suites give the field comparable numbers for how well policies transfer and how badly they forget, making benchmarks like this a common yardstick for new manipulation foundation models. ### LiDAR Category: Perception URL: https://almond.bot/glossary/lidar LiDAR (light detection and ranging) is an active sensor that measures distance by emitting laser pulses and timing their reflections, producing 2D scans or 3D point clouds of the surrounding geometry. Variants include mechanically spinning units, solid-state and MEMS-mirror designs, and flash LiDAR. Compared with cameras, LiDAR provides direct metric depth largely independent of ambient lighting, at the cost of sparsity, price, and weaker semantic content. Why it matters for physical AI: Direct, lighting-invariant geometry remains the backbone of dependable localization and obstacle avoidance for mobile robots and autonomous vehicles, complementing learned camera-based perception where metric certainty is safety-critical. ### LiDAR Odometry Category: Navigation & SLAM URL: https://almond.bot/glossary/lidar-odometry LiDAR odometry is the estimation of a robot's incremental ego-motion by registering successive LiDAR scans, typically via iterative closest point variants or feature-based matching of edge and planar points as popularized by LOAM. Modern pipelines tightly fuse an IMU for motion prediction and scan de-skewing, as in LIO-SAM and FAST-LIO, achieving low-drift trajectories in real time even during aggressive motion. Why it matters for physical AI: Reliable ego-motion estimation without GPS enables warehouse, construction, and off-road deployment, and it supplies the accurate pose priors that mapping, navigation, and mobile manipulation stacks depend on. ### Lie Group Category: Math & Kinematics URL: https://almond.bot/glossary/lie-group A Lie group is a mathematical group that is also a smooth manifold, so that group multiplication and inversion are differentiable operations. Robotics relies chiefly on SO(3), the group of 3D rotations, and SE(3), the group of rigid-body transformations, whose associated Lie algebras provide local linearizations used via the exponential and logarithm maps. This machinery underlies state estimation on manifolds, trajectory optimization, and modern kinematics formulations such as product of exponentials. Why it matters for physical AI: Correctly handling rotation geometry, rather than treating poses as flat vectors, is essential for stable state estimators, well-conditioned policy action spaces, and physically sensible interpolation of end-effector motions. ### Lifelong Learning Category: Robot Learning URL: https://almond.bot/glossary/lifelong-learning Lifelong learning, also called continual learning, is the ability of a model to acquire new tasks and knowledge sequentially over time without catastrophically forgetting previously learned capabilities. Techniques include regularization methods such as elastic weight consolidation, experience replay from stored or generated data, and modular or adapter-based architectures that isolate task-specific parameters. Benchmarks such as LIBERO evaluate these trade-offs in robot manipulation settings. Why it matters for physical AI: Deployed robots continually encounter new objects, layouts, and customer requests; models that must be retrained from scratch for every addition cannot scale, so continual adaptation is central to long-lived fleets. ### Limit Switch Category: Hardware URL: https://almond.bot/glossary/limit-switch A limit switch is an electromechanical or contactless sensor that triggers when a moving axis reaches a defined position, used to bound travel and to establish a repeatable reference during homing. Mechanical lever, optical interrupter, and Hall-effect variants are common. On robots and CNC-style axes, limit switches provide hard end-of-travel protection independent of software and let incremental-encoder systems recover absolute position after power-up. Why it matters for physical AI: Cheap, deterministic hardware interlocks remain a bedrock safety layer beneath learned controllers, guaranteeing that policy errors or estimator faults cannot drive an axis past its mechanical limits. ### Line Following Category: Navigation & SLAM URL: https://almond.bot/glossary/line-following Line following is a classic mobile robot navigation behavior in which the robot tracks a visual or magnetic guide line on the floor, historically using arrays of infrared reflectance sensors feeding a PID steering controller. It remains widely used by automated guided vehicles in factories and warehouses, where taped or painted paths provide simple, highly reliable routing, and it serves as a canonical introductory problem in robotics education. Why it matters for physical AI: Guide-line navigation illustrates the infrastructure-versus-autonomy trade-off: fixed paths deliver near-perfect reliability cheaply, setting the dependability bar that free-ranging learned navigation must meet to displace them. ### Linear Actuator Category: Hardware URL: https://almond.bot/glossary/linear-actuator A linear actuator is a device that produces motion along a straight line, implemented electrically via lead screws, ball screws, belt drives, rack-and-pinion mechanisms, or direct-drive linear motors, or via pneumatic and hydraulic cylinders. Key specifications include stroke length, force, speed, backdrivability, and positioning accuracy. In robotics they appear in grippers, lift columns for mobile manipulators, gantries, and torso or leg joints. Why it matters for physical AI: Vertical lift stages and prismatic joints determine a mobile manipulator's reachable workspace, from floor to shelf height, directly shaping which real-world fetch-and-place tasks a platform can address. ### LLM Planning Category: Foundation Models URL: https://almond.bot/glossary/llm-planning LLM planning is the use of large language models to decompose natural-language goals into sequences of executable robot skills or code. SayCan grounded language model proposals with learned affordance values, Code as Policies generated executable policy code directly, and Inner Monologue incorporated textual feedback from the environment for closed-loop replanning. The LLM typically operates over a library of pretrained low-level skills rather than emitting motor commands itself. Why it matters for physical AI: Language models contribute commonsense task knowledge that robot datasets lack, letting systems sequence long-horizon household or industrial tasks from a single instruction while learned skills handle the physical execution. ### Load Cell Category: Hardware URL: https://almond.bot/glossary/load-cell A load cell is a transducer that measures force, most commonly by sensing the minute deformation of a metal flexure with bonded strain gauges arranged in a Wheatstone bridge. Signal conditioning amplifies the microvolt-level output into a calibrated force reading. In robotics, load cells appear in gripper fingertips, joint torque sensors, multi-axis force-torque sensors at the wrist, and test rigs for actuator characterization. Why it matters for physical AI: Direct force measurement grounds contact-rich manipulation: insertion, polishing, and handover tasks need force feedback that vision cannot provide, and force signals are increasingly logged as a training modality. ### Local Planner Category: Navigation & SLAM URL: https://almond.bot/glossary/local-planner A local planner is the navigation component that converts a global route into immediate velocity commands, reactively avoiding nearby obstacles while respecting the robot's kinematic and dynamic limits. Classic algorithms include the dynamic window approach (DWA) and timed elastic bands (TEB), both standard in the ROS navigation stack, with model predictive control and learned local policies increasingly used. It typically operates on a rolling local costmap at high frequency. Why it matters for physical AI: Real-world spaces change faster than maps; the local planner is where a robot's split-second reactions to people, carts, and clutter are decided, making it decisive for safe deployment among humans. ### Localization Category: Navigation & SLAM URL: https://almond.bot/glossary/localization Localization is the problem of estimating a robot's pose within a known map or reference frame from sensor data. Probabilistic approaches dominate: particle-filter methods such as adaptive Monte Carlo localization (AMCL) match LiDAR scans against occupancy grids, while Kalman-filter variants fuse GPS, IMU, and odometry outdoors. Global localization additionally resolves the kidnapped-robot problem of initializing pose without a prior estimate. Why it matters for physical AI: Every downstream navigation and mobile manipulation decision inherits localization error; centimeter-level, drift-free pose estimates are what let robots repeatably dock, pick from shelves, and operate through long unattended shifts. ### Loco-Manipulation Category: Locomotion URL: https://almond.bot/glossary/loco-manipulation Loco-manipulation is the coordinated use of locomotion and manipulation, in which a legged or mobile robot's body motion and limb contacts jointly accomplish tasks such as opening doors, carrying loads, or pushing objects. On quadrupeds this includes using legs as manipulators or mounting an arm whose motion is coupled to the base; on humanoids it extends to whole-body control that trades balance against reach and force exertion. Why it matters for physical AI: Useful work in human environments rarely separates moving from manipulating; whole-body policies that coordinate both unlock tasks like door traversal and heavy-object transport that arm-only or base-only control cannot achieve. ### Locomotion Policy Category: Locomotion URL: https://almond.bot/glossary/locomotion-policy A locomotion policy is a learned controller, typically a neural network trained with reinforcement learning in simulation, that maps proprioceptive and sometimes exteroceptive observations to joint targets for walking, running, or climbing. Trained with domain randomization and GPU-parallel simulators such as Isaac Gym, and often distilled through teacher-student schemes, such policies have achieved robust rough-terrain quadruped and humanoid locomotion, as demonstrated on ANYmal. Why it matters for physical AI: Learned locomotion is the clearest existence proof that massively parallel simulation training transfers to real dynamics, establishing the sim-to-real playbook now being applied to manipulation and humanoid whole-body control. ### Logistics Robot Category: Applications URL: https://almond.bot/glossary/logistics-robot A logistics robot is a robot deployed in warehousing, fulfillment, or freight operations, spanning goods-to-person shelf carriers of the kind popularized by Kiva Systems (now Amazon Robotics), autonomous mobile robots for case and tote transport, robotic piece-picking arms, palletizers, and automated forklifts. These systems integrate with warehouse management software and operate continuously in semi-structured facilities alongside human workers. Why it matters for physical AI: Warehouses are the largest commercial robotics market and the leading proving ground where learned picking and navigation meet hard throughput, uptime, and safety requirements at fleet scale. ### Long-Horizon Manipulation Category: Manipulation URL: https://almond.bot/glossary/long-horizon-manipulation Long-horizon manipulation is the execution of tasks composed of many sequential, interdependent steps, such as cooking a meal or assembling furniture, where success requires every stage to succeed and errors compound over minutes of execution. Approaches include hierarchical decomposition with task planners or LLMs over skill libraries, subgoal generation, and end-to-end policies trained on long demonstrations, with recovery behaviors to correct intermediate failures. Why it matters for physical AI: Compounding error is the central obstacle to economically useful autonomy: a 95 percent-reliable skill fails most twenty-step tasks, so progress on long horizons directly gates real deployment value. ### Loop Closure Category: Navigation & SLAM URL: https://almond.bot/glossary/loop-closure Loop closure is the recognition that a robot has returned to a previously visited place, and the incorporation of that constraint to correct accumulated drift in a SLAM trajectory and map. Detection uses appearance-based methods such as bag-of-visual-words retrieval or LiDAR descriptors like Scan Context, followed by geometric verification; the resulting constraint triggers a global pose-graph optimization that redistributes error over the trajectory. Why it matters for physical AI: Without loop closure, odometry drift makes large maps unusable within minutes; robust place recognition is what allows robots to maintain consistent maps across building-scale spaces and multi-hour missions. ### LoRA Category: Foundation Models URL: https://almond.bot/glossary/lora LoRA (low-rank adaptation) is a parameter-efficient fine-tuning method, introduced by Hu et al. in 2021, that freezes a pretrained model's weights and injects small trainable low-rank matrices into selected layers, typically attention projections. Only the adapter parameters, often under one percent of the model, are updated, and they can be merged into the base weights at inference with no added latency. Why it matters for physical AI: Adapting billion-parameter vision-language-action models to a new robot, gripper, or task with modest compute and data hinges on parameter-efficient methods, which also allow swapping lightweight task adapters on shared base weights. ### Loss Function Category: Robot Learning URL: https://almond.bot/glossary/loss-function A loss function is the scalar objective that quantifies the discrepancy between a model's predictions and desired outputs, whose gradient drives parameter updates during training. Robot learning employs mean squared error or L1 losses for continuous action regression, cross-entropy for tokenized actions in models like RT-2, denoising objectives for diffusion policies, and temporal-difference losses in reinforcement learning, often combined with auxiliary representation losses. Why it matters for physical AI: The loss encodes what a policy is actually optimized to do; mismatches between a convenient training objective and true task success are a persistent source of gap between benchmark scores and real-world reliability. ### Low-Cost Robot Arm Category: Hardware URL: https://almond.bot/glossary/low-cost-robot-arm A low-cost robot arm is a manipulator built from hobby-grade servos, 3D-printed structures, and open-source designs at a price of roughly one hundred to a few thousand dollars, in contrast to industrial and collaborative arms costing tens of thousands. Notable examples include the SO-100 and SO-101, the Koch v1.1 leader-follower pair, and the ALOHA rig's ViperX/WidowX setup, all designed around teleoperated data collection for learning research. Why it matters for physical AI: Cheap arms transformed who can do manipulation research: community-scale data collection and reproducible hardware baselines emerged only after capable platforms dropped below hobbyist price points. ### LQR Category: Control URL: https://almond.bot/glossary/lqr LQR (linear quadratic regulator) is an optimal control method that, for linear dynamics and a quadratic cost on state error and control effort, yields the optimal linear state-feedback law u = -Kx, with the gain computed by solving an algebraic Riccati equation. Nonlinear robots use LQR on local linearizations, in gain scheduling, for trajectory stabilization via time-varying LQR, and inside iterative variants such as iLQR for trajectory optimization. Why it matters for physical AI: Principled feedback synthesis with provable properties remains indispensable for stabilizing balancing robots and tracking planned trajectories, and iLQR-style solvers power the model-based planning layers beneath many learned systems. ### LSTM Category: Robot Learning URL: https://almond.bot/glossary/lstm LSTM (long short-term memory) is a recurrent neural network architecture, introduced by Hochreiter and Schmidhuber in 1997, that uses gated memory cells with input, forget, and output gates to preserve information over long sequences and mitigate vanishing gradients. Before transformers became dominant, LSTMs were the standard choice for sequence modeling, and they remain in use where compact recurrent state and strictly streaming inference are advantageous. Why it matters for physical AI: Recurrent policies carry a fixed-size memory across timesteps, suiting partially observable control on compute-limited hardware, and they remain common in reinforcement-learned locomotion controllers that must run at kilohertz rates. ## Glossary: M ### Machine Vision Category: Perception URL: https://almond.bot/glossary/machine-vision Machine vision is the engineering discipline of using cameras, optics, structured lighting, and image processing to perform automated inspection, measurement, and guidance in industrial settings. Classical pipelines rely on calibrated setups and deterministic algorithms for tasks such as defect detection, barcode reading, gauging, and part localization for robot guidance, with deep learning increasingly handling variable appearance and defect classification. Why it matters for physical AI: Factory vision systems set the reliability standard, often above 99.9 percent, that learned open-world perception is measured against, and vision-guided picking remains the entry point for robot arms into production lines. ### Magnetometer Category: Hardware URL: https://almond.bot/glossary/magnetometer A magnetometer is a sensor that measures magnetic field strength and direction, used in robotics primarily to estimate absolute heading from the Earth's magnetic field as the third component of nine-axis inertial measurement units. Readings require calibration for hard-iron and soft-iron distortions, and they degrade near motors, ferrous structures, and power electronics, which limits reliability indoors and on electrically noisy platforms. Why it matters for physical AI: Drift-free heading is the hardest component of orientation estimation to obtain; magnetometers provide it cheaply outdoors, while their indoor unreliability drives the need for visual or LiDAR-based heading correction. ### Manifold Category: Math & Kinematics URL: https://almond.bot/glossary/manifold A manifold is a topological space that locally resembles Euclidean space, allowing calculus to be performed on curved spaces through local coordinate charts. Robot configuration spaces are manifolds: the orientation of a rigid body lives on SO(3), a full pose on SE(3), and a revolute joint's angle on the circle S1. Respecting manifold structure matters for interpolation, optimization, probability distributions, and filtering over robot states. Why it matters for physical AI: Treating rotations and poses as flat vectors produces singularities and biased estimators; manifold-aware optimization and learning are prerequisites for numerically sound state estimation, motion planning, and policy action parameterizations. ### Manipulability Ellipsoid Category: Math & Kinematics URL: https://almond.bot/glossary/manipulability-ellipsoid The manipulability ellipsoid is a geometric representation of a manipulator's instantaneous capacity to generate end-effector velocities in different directions, derived from the singular value decomposition of the Jacobian at a given configuration. Introduced by Yoshikawa, whose scalar manipulability measure is proportional to the ellipsoid's volume, it shrinks to lower dimension near singularities. A dual force ellipsoid characterizes achievable end-effector forces. Why it matters for physical AI: Configuration-dependent dexterity determines where in the workspace a robot can move and exert force effectively; manipulability metrics guide redundancy resolution, workcell layout, and grasp approach selection. ### Manipulation Category: Manipulation URL: https://almond.bot/glossary/manipulation Manipulation is the purposeful control of objects in the environment through selective contact, encompassing grasping, pick-and-place, insertion, tool use, and deformable object handling. It requires coordinating perception, motion planning, and contact-rich control under uncertainty about object geometry, pose, and physical properties. Despite decades of progress, general-purpose manipulation in unstructured environments remains one of the defining open problems of robotics. Why it matters for physical AI: Most economically valuable physical work is manipulation, and its combinatorial diversity of objects and contacts is precisely the regime where learned, data-driven policies are displacing hand-engineered pipelines. ### Manipulator Category: Hardware URL: https://almond.bot/glossary/manipulator A manipulator is a robotic arm consisting of a kinematic chain of links and actuated joints terminating in an end-effector such as a gripper or tool. Common configurations include six or seven degree-of-freedom articulated arms, SCARA arms for planar assembly, delta robots for high-speed picking, and Cartesian gantries. Key specifications include payload, reach, repeatability, speed, and whether the arm is force-limited for collaborative operation. Why it matters for physical AI: The manipulator is the physical substrate every learned policy must ultimately command; its kinematics, compliance, and sensing define the action space and safety envelope available to manipulation foundation models. ### Manufacturing Execution System (MES) Category: Software & Middleware URL: https://almond.bot/glossary/manufacturing-execution-system A manufacturing execution system (MES) is the software layer that manages and monitors production on the factory floor, sitting between enterprise resource planning systems and shop-floor equipment. It handles work order dispatch, routing, traceability, quality data collection, and equipment status. Robots and automated cells integrate with the MES to receive jobs, report completions and defects, and synchronize with upstream and downstream operations. Why it matters for physical AI: Robots earn their keep only when woven into production workflows; MES integration is how autonomous cells receive tasks, report outcomes, and become accountable parts of a traceable manufacturing process. ### Markov Decision Process (MDP) Category: Robot Learning URL: https://almond.bot/glossary/markov-decision-process A Markov decision process (MDP) is the standard mathematical formalism for sequential decision-making, defined by a state space, action space, transition dynamics, reward function, and discount factor, under the Markov assumption that the current state fully summarizes history. Solving an MDP means finding a policy maximizing expected cumulative reward. Reinforcement learning addresses MDPs with unknown dynamics; partially observable variants (POMDPs) relax the full-observability assumption. Why it matters for physical AI: Nearly all reinforcement learning theory and algorithms for robot control assume an underlying MDP, and understanding where robotics violates its assumptions, especially partial observability, explains many practical failures. ### Masked Autoencoder (MAE) Category: Foundation Models URL: https://almond.bot/glossary/masked-autoencoder A masked autoencoder (MAE) is a self-supervised vision pretraining method, introduced by He et al. in 2022, that masks a large fraction of image patches, typically 75 percent, and trains a vision transformer to reconstruct the missing pixels. An asymmetric design in which the encoder sees only visible patches makes pretraining efficient. The learned representations transfer strongly to downstream recognition and control tasks. Why it matters for physical AI: Masked-reconstruction pretraining on human video, as in the MVP line of work, produces visual encoders that markedly improve sample efficiency when reused as frozen or fine-tuned backbones for manipulation policies. ### Mass Estimation Category: Control URL: https://almond.bot/glossary/mass-estimation Mass estimation is the online or offline identification of the mass and related inertial parameters of a robot's links or a grasped payload from motion and force data. Classical approaches exploit the linearity of rigid-body dynamics in inertial parameters, solving least-squares problems over joint torques and accelerations; online variants use recursive estimators after grasping an unknown object to update feedforward compensation. Why it matters for physical AI: Picking up an unknown object changes the dynamics a controller must compensate; fast payload identification preserves tracking accuracy and stability, and is essential for safe handling of variable goods in logistics. ### Master-Follower Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/master-follower-teleoperation Master-follower teleoperation, also called leader-follower teleoperation, is a control scheme in which an operator moves a kinematically similar leader device and a follower robot tracks its joint positions in real time. Because mapping is joint-to-joint, it is intuitive and low-latency, requiring no inverse kinematics. ALOHA and the Koch and SO-100 arm pairs use this scheme with low-cost leader arms to collect fine bimanual manipulation demonstrations. Why it matters for physical AI: Joint-space puppeteering proved to be the highest-bandwidth affordable way to capture contact-rich demonstrations, and it underlies most of the bimanual manipulation datasets now used to train imitation policies. ### Material Handling Category: Applications URL: https://almond.bot/glossary/material-handling Material handling is the movement, storage, protection, and control of goods throughout manufacturing, warehousing, and distribution. Robotic material handling spans palletizing and depalletizing, case and tote transport by autonomous mobile robots, conveyor induction, truck unloading, and machine tending. It is among the largest application categories for industrial robots because the tasks are repetitive, physically demanding, and often ergonomically hazardous for human workers. Why it matters for physical AI: High volumes, structured goals, and measurable throughput make material handling the commercial beachhead where learned grasping and navigation first meet production-grade reliability requirements. ### Maximum Entropy Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/maximum-entropy-reinforcement-learning Maximum entropy reinforcement learning is a framework that augments the standard reward objective with a bonus for policy entropy, training agents to succeed while acting as randomly as possible. This encourages exploration, improves robustness by capturing multiple solution modes, and yields smoother optimization landscapes. Soft actor-critic (SAC), the most widely used off-policy algorithm for continuous control, is its best-known instantiation. Why it matters for physical AI: Sample-efficient, stable off-policy algorithms born from this framework made real-robot reinforcement learning practical, enabling locomotion and manipulation skills trained directly on hardware in hours rather than weeks. ### MCAP Category: Software & Middleware URL: https://almond.bot/glossary/mcap MCAP is an open-source container file format for heterogeneous, timestamped robotics log data, developed by Foxglove and adopted as the default storage backend for ROS 2 rosbag2. It stores serialized messages from multiple channels with self-contained schemas, supports chunked compression and indexed random access, and is serialization-agnostic, handling ROS messages, Protobuf, and JSON in a single append-friendly file suitable for high-rate logging. Why it matters for physical AI: Robot learning runs on logged experience; a fast, self-describing, tool-friendly log format streamlines the pipeline from fleet recordings to visualization, curation, and training datasets across heterogeneous systems. ### Mecanum Wheel Category: Hardware URL: https://almond.bot/glossary/mecanum-wheel A Mecanum wheel is a wheel fitted with passive rollers mounted at 45 degrees around its circumference, allowing a vehicle with four such wheels to translate in any direction and rotate in place by varying individual wheel speeds. Invented at the Swedish company Mecanum AB in the 1970s, it provides holonomic motion on flat, hard floors, at the cost of efficiency, payload-dependent slip, and sensitivity to floor debris. Why it matters for physical AI: Omnidirectional bases simplify mobile manipulation by decoupling base positioning from orientation, letting planners servo the chassis sideways under a reach target instead of executing multi-point turns in tight aisles. ### Medical Imaging Robot Category: Applications URL: https://almond.bot/glossary/medical-imaging-robot A medical imaging robot is a robotic system that positions or actuates imaging hardware, such as ultrasound probes, C-arm X-ray systems, or endoscopic cameras, to acquire diagnostic or intraoperative images with precision and repeatability beyond manual handling. Robotic ultrasound systems servo probe pose and contact force against patient anatomy, while robotic C-arms automate multi-angle acquisition for 3D reconstruction during surgery. Why it matters for physical AI: Imaging tasks demand gentle, force-controlled contact with humans under strict safety regulation, making them a demanding testbed for compliant control and for certifying learned components in clinical environments. ### MEMS Category: Hardware URL: https://almond.bot/glossary/mems MEMS (micro-electro-mechanical systems) are miniaturized devices that integrate mechanical elements and electronics on silicon using semiconductor fabrication, with feature sizes from micrometers to millimeters. In robotics, MEMS accelerometers and gyroscopes constitute virtually all low-cost inertial measurement units, and MEMS mirrors enable compact solid-state LiDAR. Mass production makes them cheap and small, though they exhibit more bias drift and noise than fiber-optic or ring-laser alternatives. Why it matters for physical AI: Commodity MEMS inertial sensing is what makes drones, legged robots, and consumer-priced platforms feasible; understanding its drift characteristics drives sensor fusion design in every modern state estimator. ### Meta-Learning Category: Robot Learning URL: https://almond.bot/glossary/meta-learning Meta-learning, or learning to learn, is the training of models that can adapt rapidly to new tasks from small amounts of data by exploiting structure shared across a distribution of training tasks. Approaches include optimization-based methods such as MAML, which finds initializations amenable to fast fine-tuning, metric-based few-shot learners, and in-context adaptation, where a sequence model conditions on examples without weight updates. Why it matters for physical AI: Robots face endless task variations that cannot all be pretrained; fast adaptation from a handful of demonstrations or trials is the capability that separates deployable generalists from fixed-repertoire systems. ### Meta-World Category: Data & Benchmarks URL: https://almond.bot/glossary/meta-world Meta-World is a simulated benchmark, introduced by Yu et al. in 2019, comprising 50 distinct tabletop manipulation tasks performed by a Sawyer arm in MuJoCo, designed to evaluate multi-task and meta-reinforcement learning. Its standardized evaluation protocols, including MT10, MT50, ML10, and ML45, measure both simultaneous multi-task mastery and few-shot adaptation to held-out tasks with shared observation and action spaces. Why it matters for physical AI: Shared-embodiment task suites isolate the algorithmic question of knowledge transfer across skills, providing the standard evidence base for claims about multi-task and meta-reinforcement learning methods before hardware trials. ### micro-ROS Category: Software & Middleware URL: https://almond.bot/glossary/micro-ros micro-ROS is the official adaptation of ROS 2 for microcontrollers, bringing nodes, publishers, subscribers, and services to resource-constrained devices running real-time operating systems such as FreeRTOS and Zephyr. It communicates with the wider ROS 2 graph through the DDS-XRCE protocol via an agent process on a host computer, letting embedded firmware for sensors and motor controllers participate directly in the robot's middleware. Why it matters for physical AI: Real robots split computation between application processors and microcontrollers; extending the middleware down to firmware unifies the stack, easing integration of low-level actuation and sensing with high-level autonomy. ### Microcontroller (MCU) Category: Hardware URL: https://almond.bot/glossary/microcontroller A microcontroller (MCU) is a single-chip computer integrating a processor core, memory, and peripherals such as timers, ADCs, PWM generators, and communication interfaces, designed for deterministic, low-power embedded control. In robots, MCUs run motor control loops, read sensors, and enforce safety interlocks with microsecond-level timing that general-purpose operating systems cannot guarantee, typically communicating with a host computer over CAN, EtherCAT, serial, or USB. Why it matters for physical AI: Hard real-time motor and safety loops live on microcontrollers regardless of how intelligent the high-level policy is; the MCU layer defines the reliable, deterministic interface that learned controllers command. ### Middleware Category: Software & Middleware URL: https://almond.bot/glossary/middleware Middleware is the software layer that handles communication and coordination between distributed components of a robot system, providing message passing, discovery, serialization, and quality-of-service management so that perception, planning, and control modules can interoperate. ROS 2, built atop DDS implementations, is the dominant robotics middleware; alternatives and complements include Zenoh, ZeroMQ-based stacks, and proprietary frameworks in industrial and automotive systems. Why it matters for physical AI: Middleware choices dictate latency, determinism, and composability across a robot's software; they are also where learned policy nodes plug into the same graph as drivers, estimators, and safety monitors. ### MimicGen Category: Data & Benchmarks URL: https://almond.bot/glossary/mimicgen MimicGen is a data generation system, introduced by NVIDIA researchers in 2023, that automatically synthesizes large imitation learning datasets from a small number of human demonstrations. It segments source demonstrations into object-centric manipulation phases, then adapts and replays those segments in new scenes with transformed object poses, verifying task success. The original work produced over 26,000 demonstrations from fewer than 200 human ones across diverse tasks. Why it matters for physical AI: Demonstration collection is the cost bottleneck of imitation learning; programmatic multiplication of scarce human data offers orders-of-magnitude leverage, a strategy now central to simulation-based data engines for manipulation. ### Minimum Jerk Trajectory Category: Control URL: https://almond.bot/glossary/minimum-jerk-trajectory A minimum jerk trajectory is a motion profile that minimizes the time integral of squared jerk, the third derivative of position, yielding smooth quintic-polynomial point-to-point movements with bell-shaped velocity profiles. Flash and Hogan showed in 1985 that human reaching movements closely match this profile. In robotics, minimum jerk and related smooth profiles reduce vibration, actuator wear, and tracking error, and make robot motion appear natural and predictable. Why it matters for physical AI: Smooth, human-like motion is both mechanically gentle and legible to nearby people, and such profiles remain standard for interpolating between waypoints produced by planners or learned policies. ### Mining Robot Category: Applications URL: https://almond.bot/glossary/mining-robot A mining robot is an autonomous or teleoperated machine performing extraction, haulage, drilling, inspection, or surveying in surface or underground mines. Deployed examples include autonomous haul truck fleets at large surface mines, automated long-distance trains, teleoperated loaders that keep operators away from unstable ground, and inspection platforms for mapping GPS-denied tunnels, a challenge highlighted by the DARPA Subterranean Challenge. Why it matters for physical AI: Mines combine extreme hazard, GPS denial, dust, and remoteness, driving some of the earliest full-scale autonomy deployments and stress-testing perception and localization far beyond benign indoor conditions. ### Mission Planning Category: Navigation & SLAM URL: https://almond.bot/glossary/mission-planning Mission planning is the highest layer of a robot's decision stack, translating operational goals into ordered sequences of tasks, routes, and contingencies, such as which waypoints to visit, which inspections to perform, and how to respond to failures or low battery. Implementations range from finite state machines and behavior trees to task-network planners and optimization-based schedulers, coordinating with fleet management for multi-robot operations. Why it matters for physical AI: Autonomy is judged over full shifts, not single skills; robust mission-level sequencing, recovery, and resource management convert individual capabilities into hours of unattended, economically meaningful operation. ### Mixed Precision Training Category: Robot Learning URL: https://almond.bot/glossary/mixed-precision-training Mixed precision training is a deep learning technique that performs most computation in reduced-precision formats such as FP16 or BF16 while keeping selected accumulations and a master copy of weights in FP32, roughly doubling throughput and halving memory on modern accelerators. FP16 training uses loss scaling to prevent gradient underflow, whereas BF16's wider dynamic range typically avoids the need, making it the default on recent hardware. Why it matters for physical AI: Training billion-parameter vision-language-action models on massive demonstration corpora is compute-bound; precision engineering substantially cuts cost and memory, and similar quantization thinking extends to fast on-robot inference. ### Mixture Density Network (MDN) Category: Robot Learning URL: https://almond.bot/glossary/mixture-density-network A mixture density network (MDN) is a neural network, introduced by Bishop in 1994, whose outputs parameterize a mixture distribution, typically Gaussian mixture weights, means, and variances, rather than a single point prediction. Trained by maximizing likelihood, MDNs represent multimodal conditional distributions. In robot learning they have been used as policy heads to capture the multiple valid actions demonstrators take in identical situations. Why it matters for physical AI: Averaging across valid strategies produces invalid actions, a core failure of naive behavior cloning; mixture heads were an early remedy and remain a lightweight alternative to diffusion-based action decoders. ### Mixture of Experts (MoE) Category: Foundation Models URL: https://almond.bot/glossary/mixture-of-experts A mixture of experts (MoE) is a neural architecture in which a learned router activates only a small subset of specialized expert subnetworks per input, decoupling total parameter count from per-token computation. Sparse MoE layers, popularized in large language models such as Switch Transformer and Mixtral, allow scaling capacity at roughly constant inference cost, at the price of routing instability and load-balancing complexity during training. Why it matters for physical AI: Robots need large model capacity but face hard onboard latency budgets; sparse expert architectures offer a path to capable policies whose per-step compute stays compatible with real-time control. ### MLP Category: Robot Learning URL: https://almond.bot/glossary/mlp An MLP (multilayer perceptron) is the elementary feedforward neural network architecture: stacked fully connected layers with nonlinear activations mapping fixed-size input vectors to outputs. Despite their simplicity, MLPs remain ubiquitous in robotics as policy and value networks for state-based reinforcement learning, as action and prediction heads atop transformer or convolutional backbones, and as the feedforward blocks inside every transformer layer. Why it matters for physical AI: Most reinforcement-learned locomotion controllers deployed on real legged robots are small MLPs, prized for microsecond inference and predictable behavior on embedded compute where large architectures cannot run. ### Mobile ALOHA Category: Robot Learning URL: https://almond.bot/glossary/mobile-aloha Mobile ALOHA is a low-cost mobile bimanual manipulation system from Stanford, introduced in 2024, that mounts an ALOHA dual-arm teleoperation rig on a wheeled base which the operator drives by walking while tethered to it, enabling whole-body demonstrations. Policies trained with imitation learning on 50 demonstrations per task, co-trained on static ALOHA data, performed household tasks such as cooking shrimp, calling an elevator, and wiping spills. Why it matters for physical AI: Whole-body household tasks were long considered out of reach for imitation learning; demonstrating them with tens of demonstrations on sub-$32,000 hardware reset expectations for data-driven mobile manipulation. ### Mobile Base Category: Hardware URL: https://almond.bot/glossary/mobile-base A mobile base is the wheeled or tracked chassis that provides a robot's locomotion, integrating drive motors, battery, drive electronics, and typically the navigation sensors such as LiDAR, cameras, and an IMU. Common drive geometries include differential drive, omnidirectional configurations using Mecanum or swerve modules, and car-like Ackermann steering, each trading maneuverability, floor requirements, payload, and odometry quality differently. Why it matters for physical AI: Base kinematics constrain everything above them: holonomic bases simplify whole-body planning for mobile manipulation, while battery capacity and payload determine how much sensing and compute a deployment can carry. ### Mobile Manipulation Category: Manipulation URL: https://almond.bot/glossary/mobile-manipulation Mobile manipulation is the integration of navigation and manipulation on a single platform, typically an arm or dual arms mounted on a mobile base, enabling tasks that require both moving through an environment and physically interacting with it. It introduces challenges beyond either capability alone: base-arm coordination and whole-body control, manipulation from imprecisely known base poses, and long-horizon tasks interleaving driving, perceiving, and grasping. Why it matters for physical AI: Fixed arms serve only what is within reach; mobility multiplies a manipulator's addressable work, making mobile manipulation the required form factor for household, healthcare, and flexible warehouse robotics. ### Mobile Robot Category: Hardware URL: https://almond.bot/glossary/mobile-robot A mobile robot is a robot capable of moving through its environment rather than being fixed to a single location, encompassing wheeled and tracked ground vehicles, legged robots, aerial drones, and marine platforms. Autonomy stacks for mobile robots center on the perception-localization-planning-control loop, distinguishing them from fixed manipulators whose workspace is static and precisely calibrated. Why it matters for physical AI: Mobility takes robots out of caged, structured workcells and into warehouses, hospitals, sidewalks, and homes, where dealing with uncontrolled environments and people defines the frontier of deployable autonomy. ### Model Predictive Control (MPC) Category: Control URL: https://almond.bot/glossary/model-predictive-control Model predictive control (MPC) is a feedback control strategy that, at every timestep, solves a finite-horizon optimal control problem using a dynamics model, applies the first action of the optimized sequence, and re-plans from the newly measured state. This receding-horizon scheme handles state and input constraints explicitly and anticipates future dynamics, at the cost of solving an optimization problem online, often at hundreds of hertz. Why it matters for physical AI: Receding-horizon replanning powers quadruped and humanoid balance, drone racing, and autonomous driving, and it increasingly pairs with learned dynamics models to combine data-driven accuracy with constraint-aware optimization. ### Model Predictive Path Integral (MPPI) Category: Control URL: https://almond.bot/glossary/model-predictive-path-integral Model predictive path integral (MPPI) control is a sampling-based model predictive control algorithm, developed by Williams et al., that rolls out many stochastically perturbed control sequences through a dynamics model and updates the nominal plan with an exponentially reward-weighted average. Derivative-free and trivially parallelizable on GPUs, it handles non-differentiable dynamics and costs, and was demonstrated in aggressive off-road autonomous driving on the AutoRally platform. Why it matters for physical AI: Sampling-based planning composes naturally with learned neural dynamics models and world models, since it needs only forward rollouts, making it a workhorse for model-based control with black-box simulators. ### Model-Based Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/model-based-reinforcement-learning Model-based reinforcement learning is a family of RL methods that learn an explicit model of environment dynamics and use it for policy improvement, either by generating imagined training rollouts, by planning through the model at decision time, or both. Dreamer learns latent world models and trains policies in imagination, while TD-MPC combines learned latent dynamics with short-horizon planning; both are markedly more sample-efficient than model-free baselines. Why it matters for physical AI: Real robots cannot afford millions of trial-and-error interactions; learning dynamics from limited experience and practicing in imagination is the leading strategy for sample-efficient improvement on physical hardware. ### Model-Free Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/model-free-reinforcement-learning Model-free reinforcement learning is the family of RL methods that learn policies or value functions directly from interaction without constructing an explicit dynamics model. Major branches include policy-gradient and actor-critic methods such as PPO and SAC, and value-based methods such as Q-learning. Model-free algorithms are simpler and asymptotically strong but sample-hungry, which is why robotics applies them mainly inside massively parallel simulators. Why it matters for physical AI: PPO running in GPU-parallel simulation is the engine behind nearly all deployed learned locomotion; the paradigm's sample cost is what pushed the field toward simulation-first training and sim-to-real transfer. ### Modular Robot Category: Hardware URL: https://almond.bot/glossary/modular-robot A modular robot is a robot constructed from standardized, interchangeable units, whether self-contained joint modules that let integrators compose arms of varying reach and payload, or self-reconfigurable cellular robots that can rearrange their own topology. Module interfaces standardize mechanical mounting, power, and communication. Modularity trades some performance and rigidity for reconfigurability, easier maintenance, and economies of scale in module production. Why it matters for physical AI: Cross-embodiment policies must generalize over kinematic variation, and modular hardware both generates that variation and offers a path to rapidly retargeting one trained stack across many robot configurations. ### Monocular Depth Estimation Category: Perception URL: https://almond.bot/glossary/monocular-depth-estimation Monocular depth estimation is the prediction of per-pixel scene depth from a single RGB image, an ill-posed problem solved by learning statistical regularities of scene geometry from data. Foundation models such as MiDaS and Depth Anything, trained on large mixed datasets, produce robust relative depth zero-shot, while metric variants recover absolute scale. Self-supervised training from video with photometric consistency avoids the need for ground-truth depth. Why it matters for physical AI: Extracting geometry from cheap RGB cameras removes the need for depth sensors in many manipulation and navigation stacks, and dense depth priors strengthen grasping, obstacle avoidance, and 3D scene representations. ### Monte Carlo Localization Category: Navigation & SLAM URL: https://almond.bot/glossary/monte-carlo-localization Monte Carlo localization is a particle-filter approach to robot localization, introduced by Dellaert, Fox, Burgard, and Thrun in 1999, that represents the posterior over robot pose as a set of weighted samples. Each cycle propagates particles through a motion model, reweights them by the likelihood of sensor observations against a known map, and resamples. The adaptive variant AMCL, which adjusts particle count online, is the standard localizer in the ROS navigation stack. Why it matters for physical AI: Particle filters handle the multimodal ambiguity of global localization and kidnapped-robot recovery that Gaussian filters cannot, and this algorithm still runs on most deployed indoor mobile robot fleets. ### Monte Carlo Methods Category: Math & Kinematics URL: https://almond.bot/glossary/monte-carlo-methods Monte Carlo methods are computational techniques that estimate quantities by averaging over random samples rather than computing them analytically, with accuracy improving as sample count grows. Robotics applications pervade the stack: particle filters for state estimation, sampling-based motion planners, Monte Carlo return estimates in reinforcement learning, Monte Carlo tree search for decision-making, and randomized rollouts in sampling-based controllers such as MPPI. Why it matters for physical AI: Uncertainty is irreducible in real-world sensing and dynamics; sampling-based approximations are frequently the only tractable way to propagate that uncertainty through nonlinear models for estimation, planning, and evaluation. ### Morphology Category: Hardware URL: https://almond.bot/glossary/morphology Morphology is a robot's physical form: its kinematic topology, link dimensions, actuator placement, degrees of freedom, and end-effector design. Morphology determines what behaviors are reachable and how difficult control is, a coupling studied as embodied intelligence and exploited by co-design methods that optimize body and controller jointly. In learning research, morphology is the axis of variation that cross-embodiment training seeks to generalize across. Why it matters for physical AI: Foundation models trained on datasets like Open X-Embodiment must bridge wildly different bodies; how policies represent and adapt to morphology largely determines whether one model can drive many robots. ### Motion Capture Category: Perception URL: https://almond.bot/glossary/motion-capture Motion capture is the measurement of the 3D motion of objects or bodies over time, classically with multi-camera optical systems such as Vicon and OptiTrack that triangulate retroreflective markers at hundreds of hertz with sub-millimeter accuracy. Markerless alternatives estimate pose from video using learned models. Robotics uses motion capture for ground-truth evaluation of state estimators, closed-loop control in lab settings, and recording human motion as demonstration data. Why it matters for physical AI: Precise ground truth is the reference against which localization and estimation are validated, and captured human motion feeds retargeting pipelines that supply humanoids with whole-body skill demonstrations. ### Motion Planning Category: Control URL: https://almond.bot/glossary/motion-planning Motion planning is the computation of a collision-free, kinematically feasible path for a robot from a start to a goal configuration through its configuration space. Sampling-based algorithms such as PRM, RRT, and their asymptotically optimal variants dominate high-dimensional manipulator planning, as implemented in OMPL, while optimization-based methods like CHOMP and TrajOpt refine trajectories directly. Planned paths are subsequently time-parameterized for execution. Why it matters for physical AI: Explicit planners provide the collision-safety guarantees learned policies lack, and hybrid systems increasingly pair them with learned components, using policies for contact-rich segments and planners for free-space transit. ### Motion Primitive Category: Control URL: https://almond.bot/glossary/motion-primitive A motion primitive is a parameterized, reusable unit of movement from which longer behaviors are composed, ranging from simple point-to-point segments to dynamic movement primitives (DMPs), which encode demonstrated trajectories as stable dynamical systems modulated by learned forcing terms. Primitives can be retimed, re-targeted to new goals, and sequenced or blended, providing a compact action vocabulary between raw joint commands and full task policies. Why it matters for physical AI: Structured action spaces built from primitives shrink exploration and improve safety for skill learning, and primitive libraries remain a practical middle ground where end-to-end policies are too data-hungry. ### Motion Retargeting Category: Teleoperation URL: https://almond.bot/glossary/motion-retargeting Motion retargeting is the transformation of motion from a source embodiment to a target with different kinematics, most prominently mapping human motion onto humanoid robots or animated characters. It must reconcile differing limb proportions, joint limits, and degrees of freedom while preserving the motion's intent, and for physical robots also respect balance and dynamic feasibility. Retargeted human data drives humanoid teleoperation and provides training corpora for whole-body skill learning. Why it matters for physical AI: Human movement is the natural data source for humanoid skills; high-fidelity retargeting converts motion capture and video into feasible robot trajectories, fueling teleoperation interfaces and large-scale whole-body imitation. ### Motor Controller Category: Hardware URL: https://almond.bot/glossary/motor-controller A motor controller is the electronic subsystem that drives a motor to follow commanded torque, velocity, or position, combining power electronics that switch current into the windings with feedback control running on an embedded processor. For the brushless motors common in robotics, field-oriented control (FOC) regulates torque smoothly using rotor position from an encoder, with nested velocity and position loops above and communication over buses such as CAN or EtherCAT. Why it matters for physical AI: Torque-mode motor control at kilohertz rates is the foundation for compliant, contact-safe behavior; the quality of this innermost loop bounds what any higher-level learned controller can achieve. ### MoveIt Category: Software & Middleware URL: https://almond.bot/glossary/moveit MoveIt is the standard open-source motion planning framework for ROS, providing manipulation capabilities including sampling-based planning via OMPL, collision checking with FCL, inverse kinematics through pluggable solvers, planning scene management, and trajectory execution. MoveIt 2 ports the framework to ROS 2 and adds MoveIt Servo for real-time Cartesian teleoperation. It is configured per-robot from a URDF/SRDF description via a setup assistant. Why it matters for physical AI: A shared, robot-agnostic planning stack lets teams stand up collision-aware manipulation on new arms in days, and it commonly provides the safe motion layer around learned grasp and skill policies. ### MQTT Category: Software & Middleware URL: https://almond.bot/glossary/mqtt MQTT (Message Queuing Telemetry Transport) is a lightweight publish-subscribe messaging protocol running over TCP, standardized by OASIS and ISO, in which clients exchange topic-tagged messages through a central broker with three quality-of-service levels. Designed for constrained devices and unreliable networks, it is ubiquitous in IoT and industrial telemetry, and robotics deployments commonly use it to bridge robot fleets to cloud dashboards, monitoring, and enterprise systems. Why it matters for physical AI: Fleet-scale deployment needs telemetry, health monitoring, and task dispatch across flaky wireless networks; brokered lightweight messaging complements on-robot middleware for the robot-to-cloud connectivity layer. ### MT-Opt Category: Robot Learning URL: https://almond.bot/glossary/mt-opt MT-Opt is a scalable multi-task reinforcement learning system from Google, introduced in 2021, that trained a single QT-Opt-style Q-learning policy on 12 manipulation tasks using a fleet of seven robots and a shared dataset of hundreds of thousands of real-world episodes. It exploited task interpolation, sharing data across tasks with success-detector relabeling, showing that skills learned jointly outperform independently trained ones and transfer to structurally similar new tasks. Why it matters for physical AI: Fleet-scale shared experience demonstrated concretely that data reuse across tasks accelerates real-robot reinforcement learning, prefiguring the data-pooling strategies behind today's cross-task and cross-embodiment foundation models. ### Mu-Synthesis Category: Control URL: https://almond.bot/glossary/mu-synthesis Mu-synthesis is a robust control design methodology that synthesizes controllers guaranteeing stability and performance against structured model uncertainty, quantified by the structured singular value mu. Because directly optimizing mu is intractable, practical design uses D-K iteration, alternating between H-infinity controller synthesis and computation of scaling matrices. It extends H-infinity design by exploiting the known structure of uncertainty, yielding less conservative controllers. Why it matters for physical AI: Certifiable robustness margins against quantified model error remain the gold standard for safety-critical actuation, offering the kind of formal guarantee that learned controllers are still unable to provide. ### MuJoCo Category: Simulation URL: https://almond.bot/glossary/mujoco MuJoCo (Multi-Joint dynamics with Contact) is a physics engine developed by Emo Todorov, designed for fast, stable simulation of articulated rigid bodies with contact, using generalized coordinates and a soft, convex contact model. Acquired by DeepMind in 2021 and open-sourced in 2022, it underpins benchmarks including dm_control, Meta-World, LIBERO, and Robosuite; MJX, its JAX rewrite, enables massively parallel GPU simulation. Why it matters for physical AI: Simulation speed and contact fidelity gate how much reinforcement learning experience can be generated per dollar, and this engine's balance of both made it the default substrate for continuous-control research. ### Multi-Finger Grasp Category: Manipulation URL: https://almond.bot/glossary/multi-finger-grasp A multi-finger grasp is a grasp employing three or more articulated fingers to constrain an object, enabling force-closure configurations, precision fingertip grips, and in-hand reorientation unavailable to parallel-jaw grippers. Grasp analysis characterizes such grasps through contact models and force closure, while synthesis must search a high-dimensional space of hand configurations, now increasingly handled by learned generative models trained on datasets such as DexGraspNet. Why it matters for physical AI: Parallel jaws cap the complexity of achievable manipulation; multi-finger dexterity unlocks tool use, in-hand reorientation, and human-designed object interfaces, and is advancing quickly on affordable anthropomorphic hands. ### Multi-Head Attention Category: Foundation Models URL: https://almond.bot/glossary/multi-head-attention Multi-head attention is the core operation of the transformer, introduced by Vaswani et al. in 2017, in which several attention heads run in parallel, each projecting inputs to its own query, key, and value subspaces, attending over the sequence, and concatenating results. Separate heads learn complementary relational patterns, and the mechanism's content-based, order-agnostic routing is what lets transformers fuse tokens from arbitrary modalities. Why it matters for physical AI: Attention over interleaved image patches, language tokens, and proprioceptive states is the architectural glue of vision-language-action models, letting one network bind instructions to percepts and motor outputs. ### Multi-Robot System Category: Applications URL: https://almond.bot/glossary/multi-robot-system A multi-robot system is a collection of robots that coordinate to accomplish tasks, spanning centrally orchestrated warehouse fleets, decentralized swarms, and heterogeneous teams combining ground and aerial platforms. Core problems include task allocation, often formulated as assignment or auction problems, multi-agent path finding to prevent deadlock in shared spaces, communication under bandwidth limits, and maintaining consistent shared state such as maps. Why it matters for physical AI: Commercial robotics is deployed as fleets, where hundreds of robots share aisles and job queues; coordination quality, not single-robot skill, frequently determines system throughput and becomes the operational bottleneck. ### Multi-Task Learning Category: Robot Learning URL: https://almond.bot/glossary/multi-task-learning Multi-task learning is the training of a single model on multiple tasks simultaneously so that shared structure improves data efficiency and generalization relative to per-task models. In robot learning, multi-task policies condition on task identity through one-hot vectors, goal images, or language instructions, as in RT-1 and MT-Opt. Challenges include gradient interference between tasks, data imbalance, and negative transfer, addressed by loss weighting and gradient surgery methods. Why it matters for physical AI: Generalist robot policies are, by construction, extreme multi-task learners; positive transfer across thousands of skills is the core bet behind training one foundation model instead of many specialists. ### Multi-View Geometry Category: Perception URL: https://almond.bot/glossary/multi-view-geometry Multi-view geometry is the mathematical theory relating 3D scene structure to its projections in multiple camera views, systematized in Hartley and Zisserman's standard text. Its core objects include epipolar geometry, the essential and fundamental matrices, homographies, and triangulation, with bundle adjustment jointly refining structure and camera poses. It underlies stereo vision, structure from motion, visual SLAM, and camera calibration. Why it matters for physical AI: Even as learned perception advances, geometric constraints remain the backbone of camera calibration, visual odometry, and 3D reconstruction, providing the metric scaffolding on which robot spatial reasoning is built. ### Multimodal Action Distribution Category: Robot Learning URL: https://almond.bot/glossary/multimodal-action-distribution A multimodal action distribution is a conditional distribution over actions with multiple distinct modes, arising in demonstration data whenever different experts, or the same expert at different times, solve identical situations differently, such as passing an obstacle on either side. Unimodal regression averages across modes and can output invalid intermediate actions; expressive policy heads based on mixture models, diffusion, flow matching, or autoregressive tokenization model the modes faithfully. Why it matters for physical AI: Handling demonstration multimodality is a chief reason diffusion- and tokenization-based action decoders now dominate imitation learning, directly improving success on contact-rich tasks trained from diverse human data. ### Multimodal Learning Category: Foundation Models URL: https://almond.bot/glossary/multimodal-learning Multimodal learning is the training of models that jointly process multiple input modalities, such as vision, language, audio, proprioception, and touch, by aligning them in shared representation spaces or fusing them within a single architecture. Contrastive alignment as in CLIP and unified transformer architectures over tokenized modalities are dominant patterns. Robotics extends the modality set to depth, force-torque, and tactile signals. Why it matters for physical AI: Robots inherently sense across modalities, and skills like insertion demand fusing vision with force and touch; multimodal architectures are how foundation models absorb these complementary signals into one policy. ## Glossary: N ### N-Step Return Category: Robot Learning URL: https://almond.bot/glossary/n-step-return An n-step return is a reinforcement learning target that sums n discounted rewards before bootstrapping from a value estimate at the n-th subsequent state. It interpolates between one-step temporal-difference learning (low variance, high bias) and full Monte Carlo returns (high variance, no bootstrap bias). Algorithms such as A3C and Rainbow use multi-step targets to propagate reward information faster through the value function. Why it matters for physical AI: Reward signals in robot tasks are often sparse and delayed, so multi-step targets speed up credit assignment and stabilize value learning in RL-based controllers and offline RL pipelines. ### Natural Language Instruction Category: Foundation Models URL: https://almond.bot/glossary/natural-language-instruction A natural language instruction is a free-form text command, such as 'pick up the mug and place it on the shelf', used to specify a robot's task. Language-conditioned policies embed the instruction with a pretrained language or vision-language model and condition action generation on it. Systems like SayCan, RT-2, and OpenVLA established language as a standard task interface for generalist robot policies. Why it matters for physical AI: Language provides a scalable, human-friendly way to command robots and to reuse internet-scale semantic knowledge, letting one policy cover many tasks without task-specific programming. ### Nav2 Category: Navigation & SLAM URL: https://almond.bot/glossary/nav2 Nav2 is the standard open-source navigation framework for ROS 2, succeeding the original ROS navigation stack built around move_base. It uses behavior trees to orchestrate planner, controller, smoother, and recovery servers over layered costmaps, providing global path planning and local trajectory control for mobile robots. Nav2 is widely deployed on autonomous mobile robots, service robots, and research platforms. Why it matters for physical AI: Mobile manipulation and fleet deployments need dependable point-to-point navigation; Nav2 provides a production-tested baseline so teams can focus learning efforts on manipulation and higher-level autonomy. ### NeRF (Neural Radiance Field) Category: Perception URL: https://almond.bot/glossary/nerf A NeRF (Neural Radiance Field) is a neural scene representation that maps a 3D position and viewing direction to volume density and view-dependent color, trained from posed images via differentiable volume rendering. Introduced by Mildenhall et al. in 2020, NeRFs synthesize photorealistic novel views and encode scene geometry implicitly. Variants such as Instant-NGP made training fast enough for practical robotics use. Why it matters for physical AI: Dense, differentiable 3D scene models support grasp planning, sim-to-real asset creation, and visual localization, and NeRF-style representations feed digital-twin pipelines for training manipulation policies. ### Network Latency Category: Teleoperation URL: https://almond.bot/glossary/network-latency Network latency is the time delay for data to travel between a robot and a remote operator or off-board compute, typically measured as round-trip time. In teleoperation, delays beyond roughly 100-200 milliseconds visibly degrade tracking accuracy and force-feedback stability, and jitter (variance in delay) is often more disruptive than constant lag. Latency also constrains cloud-based inference for real-time control loops. Why it matters for physical AI: Teleoperated data collection quality and the feasibility of off-board policy inference both hinge on latency budgets, shaping choices between onboard compute, edge servers, and cloud deployment. ### Neural Policy Category: Robot Learning URL: https://almond.bot/glossary/neural-policy A neural policy is a control policy represented by a neural network that maps observations, such as camera images and proprioceptive state, directly to robot actions. It is trained end-to-end with imitation learning, reinforcement learning, or both, rather than hand-engineered. Architectures range from small MLPs for joint-space control to transformer- and diffusion-based models such as ACT and Diffusion Policy. Why it matters for physical AI: Learned policies can capture contact-rich, perception-driven behaviors that are impractical to program explicitly, forming the core executable component of modern robot learning systems. ### Neural Process Category: Robot Learning URL: https://almond.bot/glossary/neural-process A neural process is a model, introduced by Garnelo et al. in 2018, that learns distributions over functions by encoding a set of context observations into a latent representation and predicting outputs, with uncertainty, at new inputs. Neural processes combine the flexibility of neural networks with Gaussian-process-like uncertainty estimates and adapt to new tasks from small context sets without gradient updates. Why it matters for physical AI: Fast adaptation from a handful of examples with calibrated uncertainty is valuable for meta-learning dynamics models and for robots that must adjust to new objects or conditions on the fly. ### Neuromorphic Computing Category: Hardware URL: https://almond.bot/glossary/neuromorphic-computing Neuromorphic computing is a hardware paradigm that mimics the brain's event-driven, spiking computation, exemplified by chips such as Intel's Loihi and IBM's TrueNorth. Instead of clocked dense matrix operations, spiking neural networks compute asynchronously and only when events occur, yielding very low power consumption. It pairs naturally with event cameras, which output sparse brightness-change events rather than frames. Why it matters for physical AI: Power budgets on mobile robots and drones are tight, and event-driven sensing plus neuromorphic inference offers microsecond-latency, milliwatt-scale perception for high-speed control. ### Newton-Euler Dynamics Category: Math & Kinematics URL: https://almond.bot/glossary/newton-euler-dynamics Newton-Euler dynamics is a formulation of rigid-body dynamics that applies Newton's second law and Euler's rotational equations to each link of a robot. The recursive Newton-Euler algorithm (RNEA) computes inverse dynamics, the joint torques required for a desired motion, in time linear in the number of joints, making it a workhorse of real-time controllers. It is the standard alternative to the Lagrangian formulation. Why it matters for physical AI: Fast, exact inverse dynamics underpins computed-torque control, whole-body controllers, and the physics engines used to train locomotion and manipulation policies in simulation. ### Next-Token Prediction Category: Foundation Models URL: https://almond.bot/glossary/next-token-prediction Next-token prediction is the training objective in which a model learns to predict the next discrete token in a sequence given all previous tokens, the recipe behind GPT-style large language models. Applied to robotics, actions, states, and sensor readings are discretized into tokens so a single autoregressive transformer can model trajectories, as in RT-1, RT-2, and Gato. Why it matters for physical AI: Casting control as sequence modeling lets robotics borrow architectures, scaling laws, and pretraining infrastructure from language modeling, a foundation of current vision-language-action models. ### Noise Injection Category: Robot Learning URL: https://almond.bot/glossary/noise-injection Noise injection is the deliberate addition of perturbations to actions, observations, or system parameters during training or data collection. In imitation learning, methods such as DART inject noise into expert demonstrations so the dataset covers states near, but off, the expert trajectory, mitigating covariate shift. In reinforcement learning and sim-to-real transfer, action and dynamics noise improve exploration and robustness. Why it matters for physical AI: Policies trained only on clean expert states fail once small errors compound; training with realistic perturbations produces controllers that recover instead of drifting further off-distribution. ### Noise Schedule Category: Robot Learning URL: https://almond.bot/glossary/noise-schedule A noise schedule is the sequence of noise levels used in a diffusion model, defining how much Gaussian noise is added at each step of the forward process and, correspondingly, how denoising proceeds during sampling. Common choices include linear and cosine schedules over the variance parameters. In diffusion policies, the schedule and the number of denoising steps trade off action quality against inference latency. Why it matters for physical AI: Diffusion-based robot policies must denoise actions within a real-time control budget, so schedule design and step-reduction techniques like DDIM directly determine achievable control rates. ### Nominal Trajectory Category: Control URL: https://almond.bot/glossary/nominal-trajectory A nominal trajectory is a reference state and control sequence around which a robot's motion is planned and stabilized. Trajectory optimizers such as iLQR or direct collocation produce a nominal plan, and feedback controllers, for example time-varying LQR, correct deviations from it during execution. The term contrasts the idealized planned motion with the perturbed trajectory the robot actually realizes. Why it matters for physical AI: Real robots never track plans exactly; separating a nominal plan from feedback stabilization is the standard structure for reliable execution of dynamic locomotion and manipulation. ### Non-prehensile Manipulation Category: Manipulation URL: https://almond.bot/glossary/non-prehensile-manipulation Non-prehensile manipulation is manipulation without grasping, moving objects by pushing, sliding, pivoting, toppling, tossing, or rolling them. Because the object is not rigidly held, its motion depends on frictional contact mechanics, studied foundationally in Matthew Mason's work on pushing. It extends what a robot can handle to objects too large, flat, or heavy to grasp. Why it matters for physical AI: Warehouses and homes are full of items that cannot be grasped directly, and learned non-prehensile skills like pushing and flipping greatly expand a manipulator's practical task coverage. ### Nonholonomic Constraint Category: Math & Kinematics URL: https://almond.bot/glossary/nonholonomic-constraint A nonholonomic constraint is a constraint on a system's velocities that cannot be integrated into a constraint on configuration alone. The rolling-without-slipping condition of wheels is the canonical example: a differential-drive robot cannot translate sideways, yet can still reach any planar pose through maneuvering. Such systems require specialized motion planners, since feasible paths must respect the velocity constraints at every point. Why it matters for physical AI: Wheeled bases, cars, and steerable platforms dominate mobile robotics, so planners and learned navigation policies must produce trajectories that respect nonholonomic kinematics to be executable. ### Nonlinear Model Predictive Control (NMPC) Category: Control URL: https://almond.bot/glossary/nonlinear-model-predictive-control Nonlinear model predictive control (NMPC) is a control strategy that repeatedly solves a finite-horizon optimal control problem with nonlinear dynamics and constraints, applies the first action, and re-plans at the next step. Solvers based on sequential quadratic programming or differential dynamic programming, and toolchains like acados, enable high-rate NMPC. It is standard in legged locomotion, agile drone flight, and mobile manipulation. Why it matters for physical AI: Dynamic robots operating near actuation and contact limits benefit from controllers that reason about constraints explicitly, and NMPC often serves as the low-level backbone beneath learned high-level policies. ### Nonlinear Optimization Category: Math & Kinematics URL: https://almond.bot/glossary/nonlinear-optimization Nonlinear optimization is the minimization of a nonlinear objective subject to nonlinear constraints, the mathematical core of trajectory optimization, inverse kinematics, calibration, and SLAM back-ends. Robotics problems are typically solved with local gradient-based methods, sequential quadratic programming or interior-point solvers such as IPOPT and SNOPT, which find locally optimal solutions given a good initial guess. Why it matters for physical AI: Most planning and estimation problems in robotics reduce to nonlinear programs, so solver speed, convergence behavior, and initialization strategies directly bound what runs on a real robot. ### Normal Estimation Category: Perception URL: https://almond.bot/glossary/normal-estimation Normal estimation is the computation of surface normal vectors from 3D data such as point clouds or depth images, typically by fitting a local plane to each point's neighborhood via principal component analysis. Estimated normals describe local surface orientation and are sensitive to sensor noise and neighborhood size. Libraries such as PCL and Open3D provide standard implementations. Why it matters for physical AI: Grasp planners align gripper approach directions with surface normals, and normals feed registration, segmentation, and contact reasoning throughout manipulation perception pipelines. ### Normalization Statistics Category: Robot Learning URL: https://almond.bot/glossary/normalization-statistics Normalization statistics are the per-dimension values, typically means and standard deviations or quantile bounds, used to scale observations and actions before training a policy and to unscale predicted actions at inference time. When training on mixed datasets such as Open X-Embodiment, statistics are often computed per dataset or per embodiment. Checkpoints must ship with the exact statistics used during training. Why it matters for physical AI: Mismatched normalization is a common silent failure when fine-tuning or deploying pretrained robot policies, producing wrongly scaled actions despite a correctly loaded model. ### Normalizing Flow Category: Robot Learning URL: https://almond.bot/glossary/normalizing-flow A normalizing flow is a generative model that transforms a simple base distribution into a complex one through a sequence of invertible, differentiable mappings, allowing exact likelihood computation via the change-of-variables formula. Architectures such as RealNVP and neural spline flows are used in robotics for density estimation, multimodal policy distributions, and anomaly detection. Why it matters for physical AI: Exact densities make flows useful for detecting out-of-distribution observations and for modeling multimodal action distributions where a Gaussian policy would average between valid behaviors. ### NPU (Neural Processing Unit) Category: Hardware URL: https://almond.bot/glossary/npu An NPU (neural processing unit) is a processor specialized for neural network inference, accelerating the matrix multiplications and convolutions that dominate deep learning workloads at low power. NPUs appear as dedicated blocks in edge systems-on-chip, including robotics-focused platforms such as the NVIDIA Jetson family (alongside its GPU and DLA accelerators) and Qualcomm's robotics processors. Why it matters for physical AI: Running perception and policy inference within a mobile robot's power and thermal envelope depends on accelerator throughput, making NPU capability a key sizing constraint for onboard deployment. ### Nuclear Robot Category: Applications URL: https://almond.bot/glossary/nuclear-robot A nuclear robot is a robot designed to operate in radioactive environments for inspection, maintenance, decommissioning, and disaster response, where human access is dangerous or impossible. Notable deployments include remotely operated manipulators at decommissioning sites such as Sellafield and the robots sent into Fukushima Daiichi after the 2011 accident. Designs emphasize radiation tolerance, sealed mechanics, and reliable teleoperation. Why it matters for physical AI: Hazardous-environment work is a proving ground for teleoperation and shared autonomy, pushing advances in remote manipulation that transfer to broader physical AI deployment. ### Null-Space Control Category: Control URL: https://almond.bot/glossary/null-space-control Null-space control is a technique for redundant manipulators that executes secondary objectives using joint motions that do not affect the primary task. Motions are projected through the null space of the task Jacobian, so a 7-DoF arm can, for example, track an end-effector pose while simultaneously avoiding joint limits, obstacles, or singular configurations. It is a core mechanism in operational space and whole-body control. Why it matters for physical AI: Redundant arms and humanoids have more joints than a task strictly needs; exploiting that redundancy safely is essential for natural, collision-free behavior around people and clutter. ### Null-Space Grasping Category: Manipulation URL: https://almond.bot/glossary/null-space-grasping Null-space grasping refers to controlling the internal forces of a multi-contact grasp, the contact force combinations lying in the null space of the grasp matrix, which squeeze the object without producing any net wrench on it. Modulating these internal forces lets a hand or dual-arm system hold an object more or less tightly and maintain friction constraints without disturbing the object's motion. Why it matters for physical AI: Multi-fingered and dual-arm holds require regulating squeeze forces independently of object motion, which is central to secure yet gentle handling of fragile or heavy items. ### Numerical Inverse Kinematics Category: Math & Kinematics URL: https://almond.bot/glossary/numerical-inverse-kinematics Numerical inverse kinematics is the iterative computation of joint angles that achieve a desired end-effector pose, typically via Jacobian-based updates such as the pseudoinverse or damped least squares, or by posing IK as a nonlinear optimization. Solvers like TRAC-IK and optimization-based formulations handle joint limits, redundancy, and secondary costs that closed-form analytic methods cannot. Why it matters for physical AI: General-purpose arms with seven or more joints rarely admit practical analytic IK, so fast, robust numerical solvers sit inside nearly every manipulation planning and teleoperation stack. ## Glossary: O ### Object Detection Category: Perception URL: https://almond.bot/glossary/object-detection Object detection is the computer vision task of localizing and classifying objects in an image, producing bounding boxes or instance masks with category labels. Modern detectors span two-stage models like Faster R-CNN, single-stage models like the YOLO family, and transformer-based approaches like DETR, with open-vocabulary variants such as OWL-ViT and Grounding DINO detecting categories described in free text. Why it matters for physical AI: Detecting and localizing task-relevant objects is the entry point of most manipulation pipelines, and open-vocabulary detectors let robots ground language instructions in what they see. ### Object Pose Tracking Category: Perception URL: https://almond.bot/glossary/object-pose-tracking Object pose tracking is the continuous estimation of an object's 6-DoF position and orientation across a video or sensor stream, updating the pose as the object or camera moves. Trackers combine an initial pose estimate with frame-to-frame updates using image features, depth alignment, or learned models such as FoundationPose, and must remain robust to occlusion and fast motion. Why it matters for physical AI: Closed-loop manipulation, such as regrasping, insertion, and handovers, depends on knowing where an object is throughout the interaction, not just at the first frame. ### Object-Centric Representation Category: Perception URL: https://almond.bot/glossary/object-centric-representation An object-centric representation is a scene encoding structured as a set of discrete object entities, each with its own features such as position, pose, extent, and appearance, rather than a single global feature vector or raw pixels. Approaches range from unsupervised slot-based models like Slot Attention to detector-derived object tokens consumed by manipulation policies. Why it matters for physical AI: Factoring scenes into objects supports compositional generalization: a policy that reasons over object slots can transfer skills across novel arrangements and distractors more readily than pixel-level policies. ### Observation History Category: Robot Learning URL: https://almond.bot/glossary/observation-history Observation history is the window of past observations, and often past actions, provided to a policy at each decision step instead of only the current frame. History lets a policy infer velocities, contact states, and other partially observed quantities, implemented via frame stacking, recurrent networks, or transformer context. Many manipulation policies use short histories, since longer windows can encourage causal-confusion shortcuts. Why it matters for physical AI: Real robots are partially observed, so temporal context is often required, but history length is a consequential design choice affecting both robustness and imitation-learning failure modes. ### Observation Space Category: Robot Learning URL: https://almond.bot/glossary/observation-space An observation space is the formal specification of what a policy perceives at each timestep: the set of sensor modalities, their shapes, ranges, and encodings, such as RGB images from specified cameras, joint positions, gripper state, and force-torque readings. It defines the input interface of a learned policy, mirroring the action space on the output side. Why it matters for physical AI: Policy transfer between simulation, datasets, and hardware requires exactly matching observation spaces, and modality choices, such as adding wrist cameras or tactile input, strongly shape learnability. ### Obstacle Avoidance Category: Navigation & SLAM URL: https://almond.bot/glossary/obstacle-avoidance Obstacle avoidance is the capability of a robot to detect and steer around objects blocking its motion, spanning reactive local methods and deliberative planning. Classical techniques include artificial potential fields, the dynamic window approach, and velocity obstacles for moving agents, while manipulation stacks avoid obstacles through collision checking in motion planners. Learned policies increasingly handle avoidance implicitly from perception. Why it matters for physical AI: Safe operation around people, clutter, and other robots is non-negotiable for deployment, making reliable avoidance a baseline requirement for both navigation and arm motion generation. ### Occlusion Category: Perception URL: https://almond.bot/glossary/occlusion Occlusion is the blocking of a camera's view of an object or region by other geometry, including self-occlusion, where a robot's own arm hides the workspace. Occlusions create missing or ambiguous data for detection, pose estimation, and tracking, and are typically mitigated with multiple viewpoints, temporal filtering, active perception, or amodal models that reason about hidden extents. Why it matters for physical AI: Manipulation constantly generates occlusion, since the gripper covers exactly what it interacts with, so policies and perception systems must remain reliable when key pixels vanish mid-task. ### Occupancy Grid Category: Navigation & SLAM URL: https://almond.bot/glossary/occupancy-grid An occupancy grid is a spatial map that discretizes the environment into cells, each storing the probability that it is occupied by an obstacle. Introduced by Moravec and Elfes in the 1980s, occupancy grids are updated from range-sensor measurements via Bayesian filtering and remain the dominant map representation for 2D navigation; OctoMap extends the idea to 3D with octrees. Why it matters for physical AI: Grid maps are the common currency between SLAM, global planners, and local controllers in fielded mobile robots, prized for simplicity, probabilistic noise handling, and constant-time lookups. ### Occupancy Map Learning Category: Perception URL: https://almond.bot/glossary/occupancy-map-learning Occupancy map learning is the use of learned models to predict which regions of space are occupied, going beyond direct sensor ray-casting to infer occupancy in unobserved or occluded areas. Examples include neural implicit occupancy fields, semantic occupancy prediction from camera images in autonomous driving, and models that forecast how occupancy will evolve around dynamic agents. Why it matters for physical AI: Predicting occupancy behind occlusions and ahead in time lets robots plan safely with incomplete sensing, a step beyond maps that only record what sensors have directly measured. ### Octo Category: Foundation Models URL: https://almond.bot/glossary/octo Octo is an open-source generalist robot policy released by a UC Berkeley-led team in 2024, trained on roughly 800,000 trajectories from the Open X-Embodiment dataset. It uses a transformer backbone with a diffusion action head, supports language and goal-image conditioning, and is designed to be fine-tuned to new robots, sensors, and action spaces with modest data and compute. Why it matters for physical AI: Octo demonstrated that mid-sized open models pretrained on cross-embodiment data can be adapted cheaply to new platforms, making it a common baseline and starting point for policy fine-tuning. ### Odometry Category: Navigation & SLAM URL: https://almond.bot/glossary/odometry Odometry is the estimation of a robot's change in pose over time from onboard motion measurements, such as wheel encoder counts, IMU integration, or visual and LiDAR feature tracking. Because each increment carries error, purely odometric estimates drift without bound and are typically fused with loop closures or global localization. Visual-inertial odometry is standard on GPS-denied aerial and legged platforms. Why it matters for physical AI: Nearly every mobile robot's state estimator is built on odometry, and its drift characteristics determine how much a system must rely on mapping, beacons, or learned corrections. ### OEE (Overall Equipment Effectiveness) Category: Applications URL: https://almond.bot/glossary/oee OEE (overall equipment effectiveness) is a manufacturing metric that multiplies three factors, availability, performance, and quality, to express what fraction of planned production time yields good parts at full speed. An OEE of 100 percent means no downtime, no slow cycles, and no defects. It is the standard yardstick for evaluating automation investments, including robotic workcells. Why it matters for physical AI: Robot deployments in factories are judged by their effect on OEE, so the uptime, cycle time, and error recovery of learned systems translate directly into this metric. ### Offline Reinforcement Learning (Offline RL) Category: Robot Learning URL: https://almond.bot/glossary/offline-reinforcement-learning Offline reinforcement learning (offline RL) is the training of policies from a fixed, previously collected dataset without further environment interaction. Its central challenge is distributional shift: naive value estimation overestimates out-of-distribution actions, which algorithms like CQL, IQL, and TD3+BC counteract with conservatism or behavior regularization. Benchmarks such as D4RL standardized evaluation in the field. Why it matters for physical AI: Robots cannot cheaply explore in the real world, so learning to improve on logged demonstrations and past deployments without new trials is a key route to scalable policy improvement. ### Offline-to-Online RL Category: Robot Learning URL: https://almond.bot/glossary/offline-to-online-reinforcement-learning Offline-to-online RL is a training regime that first pretrains a policy and value function on a static dataset with offline reinforcement learning, then continues improving them with limited online interaction. The offline phase provides a competent starting point, while the online phase corrects value errors and refines behavior; methods such as AWAC and Cal-QL address the instability that can occur at the handoff. Why it matters for physical AI: Warm-starting from logged data slashes the real-world interaction needed to reach reliable performance, a practical recipe for improving deployed robot skills with minutes rather than days of trials. ### OMPL (Open Motion Planning Library) Category: Software & Middleware URL: https://almond.bot/glossary/ompl OMPL (Open Motion Planning Library) is an open-source C++ library of sampling-based motion planning algorithms, including RRT, RRT*, PRM, and KPIECE, developed at Rice University's Kavraki Lab. OMPL deliberately excludes collision checking and robot models, operating over abstract configuration spaces, which lets frameworks like MoveIt supply the geometry and use OMPL as their default planning backend. Why it matters for physical AI: A large share of real-world arm motions are still produced by sampling-based planners, and OMPL remains the reference implementation underlying manipulation stacks that learned policies must integrate with. ### Onboard Compute Category: Hardware URL: https://almond.bot/glossary/onboard-compute Onboard compute is the processing hardware carried on the robot itself, running perception, state estimation, planning, and policy inference locally rather than on remote servers. Typical choices range from embedded systems-on-chip like the NVIDIA Jetson family to full GPUs on larger platforms, constrained by power draw, heat dissipation, and weight. It contrasts with off-board or cloud compute reached over a network. Why it matters for physical AI: Model size, control frequency, and autonomy all bottleneck on the watts and FLOPS available onboard, driving quantization, distillation, and hierarchical designs that keep large models off the critical control path. ### One-Shot Imitation Learning Category: Robot Learning URL: https://almond.bot/glossary/one-shot-imitation-learning One-shot imitation learning is the problem of executing a new task after seeing a single demonstration of it, rather than training on a large task-specific dataset. Approaches meta-train across distributions of tasks so the policy can condition on a demonstration at test time, as in the one-shot imitation work of Duan et al. and meta-learning approaches following MAML. Recent variants condition generalist policies on demonstration videos or trajectory prompts. Why it matters for physical AI: Teaching a robot a new task from one showing, instead of hundreds of teleoperated episodes, would change the economics of deployment, making it a persistent benchmark for generalist policies. ### Online Fine-Tuning Category: Robot Learning URL: https://almond.bot/glossary/online-fine-tuning Online fine-tuning is the continued training of a pretrained policy using data gathered while the policy interacts with its target environment. Improvement signals include reinforcement learning rewards, human corrections as in DAgger-style interventions, or autonomous success detection. It contrasts with offline fine-tuning on a fixed dataset and requires safeguards, because the policy is updated on the same system that is acting. Why it matters for physical AI: Pretrained generalist policies rarely reach production reliability out of the box; closing the last gap with data from the actual deployment site and embodiment is often the decisive step. ### Online Reinforcement Learning (Online RL) Category: Robot Learning URL: https://almond.bot/glossary/online-reinforcement-learning Online reinforcement learning (online RL) is the classical RL setting in which an agent improves its policy while actively interacting with the environment, learning from experience gathered through its own exploration. On-policy methods like PPO and off-policy methods like SAC both operate in this regime. On real robots, online RL is constrained by wear, safety, and reset costs, which is why much of it runs in simulation. Why it matters for physical AI: Direct trial-and-error remains the most general route to high-performance control, and sim-based online RL with domain randomization produced the standout results in legged locomotion and agile flight. ### Open X-Embodiment Category: Data & Benchmarks URL: https://almond.bot/glossary/open-x-embodiment Open X-Embodiment is a large-scale collaborative robot learning dataset, released in 2023 by Google DeepMind and more than twenty partner institutions, aggregating over one million real robot trajectories spanning 22 robot embodiments into a standardized format. The accompanying RT-X experiments showed that policies trained across embodiments outperform those trained on any single robot's data alone. Why it matters for physical AI: It established cross-embodiment pretraining as the default recipe for robot foundation models; Octo, OpenVLA, and many subsequent policies train on it, making it robotics' closest analog to web-scale corpora. ### Open-Loop Control Category: Control URL: https://almond.bot/glossary/open-loop-control Open-loop control is control that executes a precomputed command sequence without using sensor feedback to correct errors during execution. It works when the system model is accurate and disturbances are small, as in some pick-and-place motions with well-fixtured parts, but any error accumulates uncorrected. It is the counterpart of closed-loop, or feedback, control. Why it matters for physical AI: Many imitation-learned policies act nearly open-loop over action chunks, and understanding when feedback is essential versus when replay suffices shapes both controller and policy architecture choices. ### Open-Loop Grasping Category: Manipulation URL: https://almond.bot/glossary/open-loop-grasping Open-loop grasping is a grasp pipeline that perceives the scene once, computes a grasp pose, and executes the reach-and-close motion blindly without further visual or tactile feedback. It is simple and fast but fails when the object moves, the pose estimate is off, or the gripper disturbs the object during approach. Closed-loop alternatives track the object visually or react to tactile signals through contact. Why it matters for physical AI: The gap between open- and closed-loop grasp success on cluttered, moving, or deformable objects is a major argument for reactive learned policies over perceive-then-execute pipelines. ### Open-Source Robotics Category: Software & Middleware URL: https://almond.bot/glossary/open-source-robotics Open-source robotics is the development and distribution of robot software, and increasingly hardware designs, under open licenses, exemplified by ROS, Gazebo, MoveIt, PyBullet, and MuJoCo on the software side and platforms like ALOHA and LeRobot's low-cost arms on the hardware side. Open ecosystems standardize interfaces and let research results transfer between labs and into products. Why it matters for physical AI: Shared stacks and cheap reproducible hardware compound progress across the field, and most robot learning research today builds directly on open simulators, datasets, and policy codebases. ### Open-Vocabulary Manipulation Category: Manipulation URL: https://almond.bot/glossary/open-vocabulary-manipulation Open-vocabulary manipulation is the ability of a robot to manipulate objects specified by arbitrary natural language, including categories never seen in its training demonstrations, rather than a fixed set of known classes. Systems achieve this by grounding language in vision-language model features, as in CLIP-based pipelines, open-vocabulary detectors, and semantic 3D scene representations, then connecting that grounding to grasping and motion generation. Why it matters for physical AI: Homes and workplaces contain unbounded object variety, so scaling beyond closed category lists is a prerequisite for general-purpose robots, and VLM grounding is currently the main lever. ### Open3D Category: Software & Middleware URL: https://almond.bot/glossary/open3d Open3D is an open-source library for 3D data processing, initiated at Intel Labs, providing data structures and algorithms for point clouds, meshes, and RGB-D images in C++ and Python. Its capabilities include point cloud registration such as ICP, normal estimation, surface reconstruction, odometry, and visualization, plus a machine learning extension for 3D deep learning workloads. Why it matters for physical AI: Point cloud and depth processing sits under most manipulation perception stacks, and Open3D is one of the standard tools for prototyping and deploying that geometry pipeline. ### OpenArm Category: Hardware URL: https://almond.bot/glossary/openarm OpenArm is an open-source robotic arm platform designed for physical AI research, with mechanical designs, electronics, and software released publicly so labs can build, modify, and repair the hardware themselves. Like other open hardware efforts in manipulation, it targets a much lower cost than commercial cobots while supporting teleoperated data collection and learned policy deployment. Why it matters for physical AI: Affordable, reproducible arms lower the barrier to collecting large demonstration datasets and validating policies on physical hardware, broadening who can contribute to manipulation research. ### OpenCV Category: Software & Middleware URL: https://almond.bot/glossary/opencv OpenCV is the most widely used open-source computer vision library, originating at Intel in 1999, offering thousands of optimized functions across image processing, feature detection, camera calibration, stereo vision, tracking, and classical machine learning in C++ and Python. In robotics it commonly handles calibration, fiducial marker detection such as ArUco, and vision preprocessing. Why it matters for physical AI: Even deep-learning-heavy robot stacks rely on classical vision plumbing, from camera calibration to fiducial tracking, and OpenCV remains the default implementation for that layer. ### OpenVLA Category: Foundation Models URL: https://almond.bot/glossary/openvla OpenVLA is an open-source 7-billion-parameter vision-language-action model released in 2024 by a Stanford-led team, built on a Llama 2 language backbone with fused SigLIP and DINOv2 visual features and trained on about 970,000 robot episodes from Open X-Embodiment. It outputs discretized actions autoregressively, reported stronger results than the much larger closed RT-2-X on evaluated tasks, and supports efficient LoRA fine-tuning to new robots. Why it matters for physical AI: As the first widely adopted fully open VLA with released weights and training code, it became a standard base model for fine-tuning and a reference point for subsequent open robot foundation models. ### Operational Design Domain (ODD) Category: Safety & Standards URL: https://almond.bot/glossary/operational-design-domain An operational design domain (ODD) is the explicit specification of conditions under which an autonomous system is designed to function safely, covering factors such as environment type, object classes, lighting, speeds, and required infrastructure. The concept was formalized in autonomous driving through SAE J3016 and is increasingly applied to mobile robots and manipulation systems to bound claims of safe operation. Why it matters for physical AI: Declaring what a robot is not designed to handle is central to safety cases, regulatory approval, and honest evaluation of learned systems whose failure modes are hard to enumerate. ### Operational Space Control (OSC) Category: Control URL: https://almond.bot/glossary/operational-space-control Operational space control (OSC) is a control framework, formulated by Oussama Khatib in 1987, that expresses robot dynamics and control objectives directly in task space, such as end-effector position and orientation, rather than in joint space. Using the task-space inertia matrix, OSC produces dynamically consistent forces for the task while secondary objectives run in the null space, enabling compliant, force-aware manipulation. Why it matters for physical AI: Task-space compliance is the natural interface for contact-rich work and for executing learned end-effector actions, making OSC a common low-level layer beneath manipulation policies. ### Optical Flow Category: Perception URL: https://almond.bot/glossary/optical-flow Optical flow is the apparent per-pixel motion field between consecutive images, induced by camera and scene motion. Classical estimators include Lucas-Kanade for sparse features and Horn-Schunck for dense fields, while learned models such as RAFT define the modern state of the art. Flow supports visual odometry, moving-object detection, and motion-based representations for policy learning. Why it matters for physical AI: Motion cues complement static appearance: flow-based features help robots track dynamic scenes, estimate ego-motion, and, in recent work, transfer manipulation skills through motion rather than pixels. ### Optimization Landscape Category: Robot Learning URL: https://almond.bot/glossary/optimization-landscape An optimization landscape is the surface a loss or objective function defines over parameter space, whose geometry, including local minima, saddle points, flat regions, and sharp valleys, governs how difficult optimization is. In robotics the term applies to neural network training as well as trajectory optimization and policy search, where contact discontinuities famously produce rugged, gradient-hostile landscapes. Why it matters for physical AI: Contact-rich dynamics create nonsmooth objectives that defeat naive gradient methods, motivating smoothing techniques, sampling-based optimizers, and reinforcement learning formulations throughout physical AI. ### Optimus (Tesla) Category: Hardware URL: https://almond.bot/glossary/optimus Optimus is Tesla's humanoid robot program, announced in 2021, developing a bipedal general-purpose robot intended first for repetitive tasks in Tesla's own factories. The program leverages Tesla's vertically integrated actuators, batteries, and vision-based AI stack derived from its driver-assistance systems, with training that draws on human video and teleoperated demonstrations. Successive prototypes have demonstrated walking, object handling, and sorting. Why it matters for physical AI: Optimus exemplifies the bet that automotive-scale manufacturing and vision-first learning can drive humanoid costs down, shaping investment and expectations across the humanoid industry. ### Oracle Policy Category: Robot Learning URL: https://almond.bot/glossary/oracle-policy An oracle policy is a policy with access to privileged information unavailable at deployment, such as exact object poses, contact states, or simulator ground truth, used as an idealized expert. In teacher-student training, an oracle teacher trained with privileged observations in simulation supervises a student policy that uses only deployable sensors, a recipe behind landmark results in learned quadruped locomotion. Why it matters for physical AI: Privileged teachers turn hard partially observed learning problems into supervised ones, one of the most effective recipes for producing deployable sim-trained policies. ### ORB Features Category: Perception URL: https://almond.bot/glossary/orb-features ORB (Oriented FAST and Rotated BRIEF) features are fast binary local image features introduced by Rublee et al. in 2011, combining the FAST corner detector with an orientation-compensated BRIEF descriptor. ORB matches at a fraction of the computational cost of SIFT or SURF while remaining rotation invariant and patent-free, which made it the backbone feature of the ORB-SLAM family of visual SLAM systems. Why it matters for physical AI: Feature-based visual SLAM and relocalization on compute-limited robots still lean on cheap, reliable descriptors, and ORB remains the canonical choice in that regime. ### Orientation Error Category: Control URL: https://almond.bot/glossary/orientation-error Orientation error is a measure of the rotational discrepancy between a desired and an actual orientation, used as the feedback signal in attitude and end-effector pose control. Common parameterizations include the angle of the relative rotation (geodesic distance on SO(3)), quaternion error, and axis-angle vectors; naive per-axis Euler-angle differences are avoided because of singularities and non-uniqueness. Why it matters for physical AI: How rotation error is represented affects controller stability and learned-policy loss design; poorly chosen orientation losses are a recurring source of subtle failures in pose tracking and imitation. ### Out-of-Distribution (OOD) Category: Robot Learning URL: https://almond.bot/glossary/out-of-distribution Out-of-distribution (OOD) describes inputs or situations that differ statistically from a model's training data, such as novel objects, lighting, backgrounds, or states reached after compounding errors. Model behavior on OOD inputs is unreliable and often confidently wrong, motivating OOD detection, uncertainty estimation, and data collection strategies that deliberately widen coverage. Why it matters for physical AI: The physical world guarantees encounters outside any training set, so detecting distribution shift and failing safely, rather than acting confidently on garbage, is core to deployable robot autonomy. ### Output Feedback Control Category: Control URL: https://almond.bot/glossary/output-feedback-control Output feedback control is control that uses only measured outputs rather than the full internal state of the system, which is often not directly observable. Designs typically combine a state estimator, such as a Kalman filter or Luenberger observer, with state feedback, or synthesize controllers directly from outputs. It contrasts with full-state feedback, which assumes every state variable is measurable. Why it matters for physical AI: Real robots never measure their full state; the observer-plus-controller pattern, and its learned analogs with latent state estimators, is how partial sensing becomes stable closed-loop behavior. ### Overfitting Category: Robot Learning URL: https://almond.bot/glossary/overfitting Overfitting is the phenomenon in which a model fits its training data, including noise and incidental correlations, so closely that performance degrades on new data. In robot learning it appears as policies that succeed on training objects, positions, and lighting but fail under small variations, or that latch onto spurious cues such as backgrounds. Remedies include more diverse data, augmentation, regularization, and early stopping. Why it matters for physical AI: Robot datasets are small and homogeneous compared with web corpora, making overfitting to a lab's specific scenes one of the most common reasons policies fail to transfer to the field. ### Overhead Camera Category: Hardware URL: https://almond.bot/glossary/overhead-camera An overhead camera is a camera mounted above the workspace looking down, providing a stable global view of the scene in manipulation setups. Overhead views simplify calibration to a work surface and offer largely occlusion-free observation of object layout, and they are commonly paired with wrist-mounted cameras that supply close-range detail during grasping. Many imitation learning rigs and benchmarks standardize on this combination. Why it matters for physical AI: Camera placement is a quiet but decisive design variable in policy learning; overhead-plus-wrist configurations dominate manipulation data collection because they balance scene context and precision. ## Glossary: P ### Packing Category: Applications URL: https://almond.bot/glossary/packing Packing is the robotic task of arranging objects into boxes, totes, or containers so they fit densely and arrive undamaged. It combines grasp planning, pose estimation, and placement optimization, and its underlying combinatorial problem, bin packing, is NP-hard, so practical systems rely on heuristics or learned policies. High SKU variety, deformable packaging, and tight container tolerances make e-commerce packing considerably harder than structured, repetitive palletizing. Why it matters for physical AI: Order fulfillment is one of the largest commercial drivers of manipulation research, and packing unseen items densely remains a stress test of how well learned policies reason about geometry, contact, and placement order. ### Painting Robot Category: Applications URL: https://almond.bot/glossary/painting-robot A painting robot is an industrial manipulator that applies paint or other coatings to parts with a spray applicator mounted as its end effector. Deployed at scale in automotive body shops since the 1980s, painting robots execute smooth, high-speed trajectories with tightly controlled standoff distance and flow rate to achieve uniform film thickness. Most are explosion-protected designs certified for operation inside flammable spray-booth atmospheres. Why it matters for physical AI: Spray painting is a canonical example of trajectory optimization under process constraints, where coverage, film uniformity, cycle time, and material waste must be traded off along a continuous tool path. ### Palletizing Robot Category: Applications URL: https://almond.bot/glossary/palletizing-robot A palletizing robot is a manipulator that stacks cases, bags, or trays onto pallets in stable, space-efficient patterns at the end of a production or packaging line. Traditional cells use four-axis arms executing pre-programmed layer patterns for known cartons, while modern mixed-case palletizing adds 3D perception and online placement planning to handle heterogeneous parcels. Payloads range from a few kilograms to several hundred. Why it matters for physical AI: Palletizing shows how perception and placement planning turn a fixed automation task into a flexible one: mixed-SKU pallets demand real-time reasoning about stability, crush resistance, and load distribution. ### PaLM-E Category: Foundation Models URL: https://almond.bot/glossary/palm-e PaLM-E is an embodied multimodal language model introduced by Google in 2023 that injects continuous sensor modalities, such as images and state estimates, directly into the token stream of the PaLM language model. Its largest variant reached 562 billion parameters and handled robot task planning, visual question answering, and captioning within a single model. PaLM-E later served as one of the vision-language backbones behind the RT-2 vision-language-action model. Why it matters for physical AI: PaLM-E was an early demonstration that web-scale vision-language pre-training transfers to embodied decision-making, shaping the now-standard recipe of building robot policies on top of pretrained multimodal backbones. ### Panoptic Segmentation Category: Perception URL: https://almond.bot/glossary/panoptic-segmentation Panoptic segmentation is a computer vision task that assigns every pixel in an image both a semantic class and, for countable objects, an instance identity, unifying semantic segmentation of background regions with instance segmentation of discrete objects. Formalized by Kirillov et al. in 2019 along with the panoptic quality metric, it produces a single coherent scene decomposition rather than two separate outputs. Why it matters for physical AI: Robots acting in clutter need to know both what surfaces are (floor, table, wall) and which pixels belong to which manipulable object, making panoptic output a natural scene representation for manipulation and navigation. ### Parallel Simulation Category: Simulation URL: https://almond.bot/glossary/parallel-simulation Parallel simulation is the practice of running thousands of physics environments simultaneously, typically on a single GPU, so that reinforcement learning agents can gather experience orders of magnitude faster than real time. NVIDIA's Isaac Gym demonstrated an end-to-end GPU pipeline in which simulation, observation processing, and policy updates all stay on-device, reducing training times for locomotion and manipulation policies from weeks to minutes or hours. Why it matters for physical AI: Massively parallel rollouts made sim-trained reinforcement learning practical for real robots: policies for quadrupeds and humanoids are now routinely trained across thousands of randomized environments before sim-to-real transfer. ### Parallel-Jaw Gripper Category: Hardware URL: https://almond.bot/glossary/parallel-jaw-gripper A parallel-jaw gripper is a two-fingered end effector whose jaws translate along a shared axis while remaining parallel to each other, closing on an object from opposite sides. Its single actuated degree of freedom makes it simple, robust, and easy to model, which is why it dominates industrial manipulation and serves as the default embodiment in most grasping datasets and robot learning benchmarks. Why it matters for physical AI: Most large-scale robot demonstration datasets and learned grasp policies assume a parallel-jaw embodiment, so its capabilities and limits, such as no in-hand reorientation, directly shape what current manipulation models can do. ### Partial Observability Category: Robot Learning URL: https://almond.bot/glossary/partial-observability Partial observability is the condition in which a robot's sensors do not reveal the full state of the environment, formalized by the partially observable Markov decision process (POMDP). Occlusions, limited fields of view, noisy sensing, and hidden object properties such as mass or friction all induce it. Agents cope by maintaining belief distributions, filtering state estimates, or conditioning policies on observation histories via recurrence or attention. Why it matters for physical AI: Real-world robotics is almost never fully observable: contact forces, occluded geometry, and object dynamics must be inferred over time, which is why memory-equipped policies often outperform purely reactive ones. ### Particle Filter Category: Navigation & SLAM URL: https://almond.bot/glossary/particle-filter A particle filter is a sequential Monte Carlo algorithm that represents the probability distribution over a robot's state with a weighted set of samples, propagating them through a motion model and reweighting them against sensor observations. Unlike Kalman filters, it handles nonlinear dynamics and multimodal beliefs, which made it the foundation of Monte Carlo Localization and the FastSLAM family of mapping algorithms. Why it matters for physical AI: Particle filters remain a workhorse for global localization, where a robot must recover its pose from complete uncertainty, and they exemplify the belief-space reasoning needed whenever perception is ambiguous. ### Passive Dynamics Category: Locomotion URL: https://almond.bot/glossary/passive-dynamics Passive dynamics refers to the motion a mechanism exhibits under gravity and its own inertia, without actuation or control. Tad McGeer's passive dynamic walkers of 1990 descended shallow slopes with a strikingly human-like gait using no motors at all, demonstrating that morphology alone can generate stable locomotion. Designers exploit passive dynamics through careful mass distribution, compliant elements, and natural oscillation frequencies to reduce actuation effort. Why it matters for physical AI: Exploiting a body's natural dynamics rather than fighting them is central to energy-efficient legged locomotion, and it motivates hardware-software co-design in which controllers work with the mechanism's intrinsic behavior. ### Passivity-Based Control Category: Control URL: https://almond.bot/glossary/passivity-based-control Passivity-based control is a design methodology that shapes a system's energy storage and dissipation so the closed loop remains passive, meaning it never outputs more energy than it has received. Because interconnections of passive systems are guaranteed stable, the approach yields controllers that remain well-behaved when a robot contacts unknown environments or humans, and it underpins stable bilateral teleoperation and interaction control. Why it matters for physical AI: Contact with unmodeled environments is where robots most often become unstable; passivity provides a principled robustness guarantee that holds regardless of what the robot touches, which is invaluable for physical human-robot interaction. ### Path Planning Category: Navigation & SLAM URL: https://almond.bot/glossary/path-planning Path planning is the problem of computing a collision-free geometric path from a start to a goal configuration through a robot's configuration space. Classical approaches include grid-based graph search such as A*, sampling-based planners such as probabilistic roadmaps and rapidly-exploring random trees, and optimization-based methods. It is distinguished from trajectory planning, which additionally assigns timing, velocities, and dynamic feasibility to the path. Why it matters for physical AI: Even as end-to-end learned policies advance, explicit planners remain the backbone of mobile robot navigation and manipulator motion generation, and hybrid systems increasingly use learned models to guide or warm-start classical planners. ### Payload Category: Hardware URL: https://almond.bot/glossary/payload Payload is the maximum mass a robot can carry while meeting its rated speed, accuracy, and duty-cycle specifications, measured at the end effector for manipulators or as transported load for mobile robots. The quoted figure must include the gripper's own mass, and manufacturers also specify allowable moments and inertias, since an offset load stresses joints more than the same mass held close to the wrist. Why it matters for physical AI: Payload budgets constrain everything from gripper and camera selection to which tasks a platform can address, and the payload-to-weight ratio is a key figure of merit for mobile manipulators and humanoids. ### Peg-in-Hole Category: Manipulation URL: https://almond.bot/glossary/peg-in-hole Peg-in-hole is the canonical contact-rich assembly task in which a robot inserts a peg into a close-fitting hole, often with clearances below a millimeter. Studied since the 1980s, it motivated the remote center of compliance device and much of the theory of fine motion under uncertainty, and it remains the standard testbed for force control, compliant insertion strategies, and learned assembly policies. Why it matters for physical AI: Insertion tolerances routinely fall below a robot's positioning accuracy, so success depends on exploiting contact feedback rather than precision alone, making peg-in-hole the benchmark for whether a method truly handles contact-rich manipulation. ### Perceiver Category: Foundation Models URL: https://almond.bot/glossary/perceiver Perceiver is a transformer architecture introduced by DeepMind in 2021 that scales to very high-dimensional inputs by cross-attending from a small set of learned latent vectors to the raw input array, decoupling compute from input size. Its successor, Perceiver IO, added flexible structured outputs. In robotics, PerAct used a Perceiver backbone to map voxelized RGB-D observations and language goals to 6-DoF manipulation actions. Why it matters for physical AI: Robot observations mix cameras, depth, proprioception, and language at very different resolutions; latent-bottleneck architectures like Perceiver offer a clean way to fuse such heterogeneous inputs without quadratic attention cost. ### Pharmaceutical Robot Category: Applications URL: https://almond.bot/glossary/pharmaceutical-robot A pharmaceutical robot is a robotic system used in drug discovery, manufacturing, or dispensing, spanning laboratory automation for high-throughput screening, aseptic filling lines, sterile compounding cells, and hospital pharmacy dispensing systems. These systems operate under strict regulatory regimes such as GMP, and cleanroom-compatible designs must minimize particle generation and tolerate frequent decontamination with harsh agents. Why it matters for physical AI: Pharma automation illustrates deployment under hard regulatory and contamination constraints, where traceability, validation, and sterility drive system design as much as speed or dexterity do. ### Photometric Calibration Category: Perception URL: https://almond.bot/glossary/photometric-calibration Photometric calibration is the process of estimating a camera's radiometric response function, vignetting profile, and exposure parameters so that recorded pixel intensities can be related to actual scene radiance. It complements geometric calibration and is essential for direct visual odometry and SLAM methods, such as DSO, that minimize photometric error between frames rather than matching sparse features. Why it matters for physical AI: Direct visual SLAM and neural scene reconstruction both assume intensities are comparable across frames; uncorrected auto-exposure and vignetting silently degrade them, so photometric calibration often decides whether such pipelines work outside the lab. ### Photorealistic Rendering Category: Simulation URL: https://almond.bot/glossary/photorealistic-rendering Photorealistic rendering is the synthesis of images that closely match real camera output, using physically based ray or path tracing, measured materials, and accurate lighting and sensor models. In robotics it is used to generate synthetic training data and to make simulators such as Isaac Sim visually faithful, narrowing the appearance gap that separates policies trained in simulation from real-world imagery. Why it matters for physical AI: The visual sim-to-real gap is often the dominant failure mode for camera-based policies; higher-fidelity rendering, combined with domain randomization, reduces how much real data is needed to close it. ### Physics Engine Category: Simulation URL: https://almond.bot/glossary/physics-engine A physics engine is software that numerically integrates the equations of motion for simulated bodies, resolving rigid-body dynamics, contact, friction, and increasingly soft-body and fluid effects. Engines widely used in robotics include MuJoCo, Bullet, NVIDIA PhysX, and Drake, which differ in contact models, solver accuracy, speed, and differentiability. Contact modeling fidelity is the key trade-off, since hard contact is mathematically stiff and expensive to resolve exactly. Why it matters for physical AI: Every simulation-trained policy inherits the biases of its physics engine, and discrepancies in contact and friction modeling are a principal cause of sim-to-real failure in manipulation and locomotion. ### Pi0 (Physical Intelligence) Category: Foundation Models URL: https://almond.bot/glossary/pi0 Pi0 is a vision-language-action foundation model released by Physical Intelligence in 2024 that pairs a pretrained vision-language backbone with a flow-matching action expert generating continuous action chunks at up to 50 Hz. Trained on a large cross-embodiment dataset spanning single-arm, dual-arm, and mobile manipulators, it demonstrated dexterous long-horizon tasks such as laundry folding and table bussing, and its open release made it a widely used base model for fine-tuning. Why it matters for physical AI: Pi0 established flow matching over continuous actions as a leading recipe for dexterous, high-frequency control, and its open weights turned pretrain-then-fine-tune into a practical workflow for real robot deployments. ### Pi0.5 (Physical Intelligence) Category: Foundation Models URL: https://almond.bot/glossary/pi0-5 Pi0.5 is the successor to Physical Intelligence's Pi0 model, introduced in 2025 to target open-world generalization. It co-trains a single vision-language-action model on heterogeneous sources, including multi-robot demonstrations, web-scale vision-language data, verbal instructions, and high-level subtask prediction, so the model both plans what to do and executes how to do it. It was demonstrated on mobile manipulators performing multi-minute cleaning tasks in homes unseen during training. Why it matters for physical AI: Generalizing to genuinely new environments, not just new objects on a familiar tabletop, is the central open problem for robot foundation models, and Pi0.5's heterogeneous co-training is one of the most prominent attacks on it. ### Pick and Place Category: Manipulation URL: https://almond.bot/glossary/pick-and-place Pick and place is the manipulation task of grasping an object at one location and setting it down at another, forming the core primitive of sorting, kitting, machine tending, and order fulfillment. Industrial variants range from high-speed delta robots moving hundreds of items per minute to perception-driven bin picking of unstructured parts, combining object detection, grasp selection, collision-free transfer motions, and controlled release. Why it matters for physical AI: Pick and place is the most economically significant manipulation primitive and the usual first benchmark for new grasping methods, yet reliable handling of arbitrary novel objects at human speed remains unsolved. ### PID Controller Category: Control URL: https://almond.bot/glossary/pid-controller A PID controller is a feedback controller that computes its output from three terms on the tracking error: a proportional term for present error, an integral term that eliminates steady-state offset, and a derivative term that damps the response. It is the most widely deployed controller in industry and closes the joint-level servo loops of virtually every robot, typically running at kilohertz rates beneath higher-level planners and policies. Why it matters for physical AI: Learned policies almost never command motor currents directly; they output setpoints tracked by PID-style loops, so the bandwidth and tuning of these inner loops bound what any higher-level controller can achieve. ### Pinch Grasp Category: Manipulation URL: https://almond.bot/glossary/pinch-grasp A pinch grasp is a precision grasp in which an object is held between fingertip surfaces, typically the thumb opposing one or more fingers, rather than enclosed by the palm and fingers as in a power grasp. It affords fine control and in-hand adjustment at the cost of strength and stability, and it is the grasp type parallel-jaw grippers approximate for small and thin objects. Why it matters for physical AI: Grasp taxonomy directly informs gripper design and policy training: tasks such as picking cards, cables, or cloth edges demand pinch precision, which drives requirements for fingertip geometry, force resolution, and tactile sensing. ### Place Recognition Category: Navigation & SLAM URL: https://almond.bot/glossary/place-recognition Place recognition is the task of identifying whether a robot's current sensor observation corresponds to a previously visited location, independent of viewpoint, lighting, or seasonal change. Visual methods evolved from bag-of-visual-words retrieval, as in FAB-MAP and DBoW, to learned global descriptors such as NetVLAD, while LiDAR approaches like Scan Context match structural signatures of 3D scans. Why it matters for physical AI: Reliable loop-closure detection is what keeps large-scale SLAM maps globally consistent, and robustness to appearance change determines whether a robot can relocalize across days, seasons, and lighting conditions. ### Placement Precision Category: Manipulation URL: https://almond.bot/glossary/placement-precision Placement precision is the accuracy with which a robot sets an object at a target pose, encompassing both position and orientation error at release. Unlike robot repeatability, it compounds uncertainty from perception, calibration, grasp acquisition, in-grasp slip, and release dynamics, so a highly repeatable arm can still place objects poorly. Tight-tolerance placement often requires in-hand pose estimation or contact-based alignment strategies. Why it matters for physical AI: Assembly, kitting, and packing are judged by where objects end up, not where the gripper went, so quantifying and improving placement error is often more important than the arm's nominal repeatability. ### Planetary Gear Category: Hardware URL: https://almond.bot/glossary/planetary-gear A planetary gear is a gear train in which several planet gears rotate around a central sun gear while meshing with an outer ring gear, with a carrier linking the planets. The arrangement delivers high torque density, coaxial input and output, and load sharing across multiple teeth. Single-stage reductions of roughly 3:1 to 10:1 with modest backlash make planetary stages common in robot joints and quasi-direct-drive actuators. Why it matters for physical AI: Low-ratio planetary reductions enable the quasi-direct-drive actuation behind modern dynamic legged robots, preserving backdrivability and torque transparency that high-ratio harmonic drives sacrifice. ### Pneumatic Actuator Category: Hardware URL: https://almond.bot/glossary/pneumatic-actuator A pneumatic actuator converts compressed air into mechanical motion, most commonly as linear cylinders, rotary vanes, or contracting McKibben artificial muscles. Pneumatics offer high force-to-weight ratio, intrinsic compliance, and low cost, but air compressibility makes precise position and force control difficult, so they traditionally serve binary open-close roles such as gripper jaws and part clamps, and they remain popular in soft robotics. Why it matters for physical AI: Intrinsic compliance makes pneumatics attractive for safe contact and soft grippers, but their control difficulty and tethered air supply explain why electric actuation dominates learned dexterous manipulation platforms. ### Point Cloud Category: Perception URL: https://almond.bot/glossary/point-cloud A point cloud is a set of 3D points, optionally carrying color, intensity, or normals, produced by LiDAR, RGB-D cameras, stereo matching, or structure-from-motion. Point clouds are unordered and vary in density with distance and viewpoint, so processing them requires permutation-invariant methods, from classical voxel grids and normal estimation to learned architectures such as PointNet and sparse 3D convolutions. Why it matters for physical AI: Point clouds are the most direct geometric input for grasp synthesis, obstacle avoidance, and mapping, and point-cloud-conditioned policies can generalize across visual appearance changes that defeat RGB-only models. ### Point Cloud Library (PCL) Category: Software & Middleware URL: https://almond.bot/glossary/point-cloud-library The Point Cloud Library (PCL) is an open-source C++ library for 2D and 3D point cloud processing, introduced by Rusu and Cousins in 2011. It provides filtering, feature estimation, keypoints, registration including ICP variants, segmentation, surface reconstruction, and model fitting with RANSAC, along with standard data structures such as k-d trees and octrees, and it integrates tightly with ROS perception pipelines. Why it matters for physical AI: PCL remains the standard toolbox for classical 3D perception in deployed systems, handling the filtering, segmentation, and registration steps that still surround learned models in most production pipelines. ### Point Cloud Registration Category: Perception URL: https://almond.bot/glossary/point-cloud-registration Point cloud registration is the problem of estimating the rigid transformation that aligns two or more point clouds of the same scene or object. The iterative closest point algorithm of Besl and McKay (1992) refines an initial guess by alternating correspondence and transform estimation, while global methods use feature matching with robust estimation, and learned approaches regress correspondences or transforms directly. Why it matters for physical AI: Registration underlies LiDAR odometry, map merging, and model-based object pose estimation for grasping, and its failure modes, such as poor initialization and partial overlap, drive much of the engineering in 3D perception stacks. ### Policy Category: Robot Learning URL: https://almond.bot/glossary/policy A policy is the function by which a robot selects actions, mapping states or raw observations to motor commands, either deterministically or as a probability distribution. In modern robot learning, policies are typically neural networks trained by reinforcement or imitation learning, consuming images and proprioception and emitting joint targets, end-effector twists, or action chunks at rates from a few hertz to several hundred. Why it matters for physical AI: The policy is the deployable artifact of robot learning: its input modalities, action space, and inference rate define the interface between learned intelligence and physical hardware. ### Policy Distillation Category: Robot Learning URL: https://almond.bot/glossary/policy-distillation Policy distillation is the transfer of behavior from one or more teacher policies into a student network by supervised learning on the teachers' outputs, introduced for deep RL by Rusu et al. in 2016. In robotics it most often takes the teacher-student form: a teacher trained in simulation with privileged state supervises a student that acts only on deployable observations such as onboard cameras and proprioception. Why it matters for physical AI: Teacher-student distillation is the standard bridge from privileged simulation training to real-world deployment, powering learned quadruped locomotion and many sim-to-real manipulation systems. ### Policy Gradient Category: Robot Learning URL: https://almond.bot/glossary/policy-gradient Policy gradient methods are reinforcement learning algorithms that directly optimize expected return by ascending its gradient with respect to policy parameters, following the REINFORCE estimator of Williams (1992) and its actor-critic refinements. Because they optimize stochastic policies over continuous action spaces without discretization, they suit robot control, with trust-region variants such as TRPO and PPO taming the high variance of the gradient estimates. Why it matters for physical AI: Continuous, high-dimensional action spaces make policy gradients the default family for training locomotion and manipulation controllers in simulation, where massive parallelism offsets their sample inefficiency. ### Policy Head Category: Foundation Models URL: https://almond.bot/glossary/policy-head A policy head is the output module of a learned control architecture that converts a shared backbone representation into robot actions. Designs include regression heads emitting continuous commands, tokenization heads producing discretized action bins as in RT-2, and generative heads such as diffusion or flow-matching experts, as in Pi0, that sample continuous action chunks. The choice governs how well multimodal action distributions and high-frequency control are handled. Why it matters for physical AI: Much of the recent progress in vision-language-action models has come from better action heads: generative chunking heads capture multimodal demonstrations and deliver smooth high-rate control where naive regression collapses to averages. ### Policy Learning Category: Robot Learning URL: https://almond.bot/glossary/policy-learning Policy learning is the umbrella term for acquiring robot control policies from data rather than hand-engineering them, spanning reinforcement learning from trial and error, imitation learning from demonstrations, and offline learning from logged experience. It contrasts with classical model-based control in optimizing behavior end to end from sensors to actions, though hybrid systems that learn components within engineered stacks are common. Why it matters for physical AI: Learning policies from data is the enabling idea of physical AI: it replaces task-by-task engineering with data collection and training, making capability scale with datasets and compute rather than expert time. ### Policy Rollout Category: Robot Learning URL: https://almond.bot/glossary/policy-rollout A policy rollout is the execution of a policy from an initial state for an episode or fixed horizon, recording the resulting trajectory of observations, actions, and rewards. Rollouts serve two roles: generating training experience for reinforcement learning, and evaluating a trained policy's success rate. Simulated rollouts are cheap and parallelizable, while real-robot rollouts are slow, risky, and require resets, motivating careful evaluation protocols. Why it matters for physical AI: Real-world evaluation bandwidth is a genuine bottleneck for robot foundation models: statistically meaningful success rates demand many physical rollouts, so rollout-efficient evaluation and trustworthy sim proxies are active research problems. ### Port-Hamiltonian Control Category: Control URL: https://almond.bot/glossary/port-hamiltonian-control Port-Hamiltonian control is an energy-based framework that models a robot as a network of energy-storing and energy-dissipating elements exchanging power through ports, with dynamics expressed via a Hamiltonian energy function. Control by interconnection and energy shaping, notably the IDA-PBC methodology, assigns a desired energy landscape and damping to the closed loop, generalizing passivity-based design with explicit physical structure. Why it matters for physical AI: Energy-based formulations yield interaction controllers with stability guarantees rooted in physics rather than linearization, and they offer a principled way to certify learned or adaptive components in contact-rich control loops. ### Pose Estimation Category: Perception URL: https://almond.bot/glossary/pose-estimation Pose estimation is the determination of an object's or robot's position and orientation, typically the full 6-DoF pose, from sensor data. Classical pipelines match features or fit models using PnP, ICP, or template matching, while learned methods regress poses directly or via keypoints, with systems such as PoseCNN and category-level and model-free approaches like FoundationPose extending to novel objects. The term also covers human and robot body pose estimation. Why it matters for physical AI: Grasping, insertion, and tool use all hinge on knowing object poses within task tolerances, and pose error is among the most common root causes of manipulation failure in deployed systems. ### Position Control Category: Control URL: https://almond.bot/glossary/position-control Position control is the actuation mode in which a robot tracks commanded joint or Cartesian positions, with servo loops rejecting disturbances to hold the setpoint regardless of required force, up to actuator limits. It is the default mode of industrial arms and excels in free-space motion, but its high stiffness makes unmodeled contact hazardous, which motivates force, impedance, or admittance control for interaction tasks. Why it matters for physical AI: Most learned policies today output position or velocity setpoints consumed by stiff industrial-style loops, so understanding position control's contact brittleness explains why compliant control interfaces matter for contact-rich learned skills. ### Positional Encoding Category: Foundation Models URL: https://almond.bot/glossary/positional-encoding Positional encoding is the mechanism by which transformers, whose attention operation is inherently permutation-invariant, receive information about token order or spatial location. The original sinusoidal scheme of Vaswani et al. (2017) has been joined by learned embeddings and relative schemes such as rotary position embeddings (RoPE). Related Fourier-feature encodings let coordinate networks like NeRF represent high-frequency spatial detail. Why it matters for physical AI: Robot policies built on transformers must encode temporal order of observations, spatial structure of image patches, and action-chunk indices, and encoding choices affect how models generalize across sequence lengths and camera configurations. ### Post-Grasp Manipulation Category: Manipulation URL: https://almond.bot/glossary/post-grasp-manipulation Post-grasp manipulation encompasses everything a robot does with an object after acquiring a grasp, including transport under dynamic constraints, in-hand reorientation, regrasping, tool use, and controlled placement or insertion. Because the initial grasp determines what downstream motions are feasible, grasp selection and post-grasp planning are coupled, a core insight formalized in task-oriented grasping and task and motion planning. Why it matters for physical AI: Picking objects up is rarely the goal; using them is. Evaluating grasps by downstream task success rather than lift stability is what separates task-driven manipulation from pick-rate benchmarks. ### Potential Field Category: Navigation & SLAM URL: https://almond.bot/glossary/potential-field A potential field is a motion generation method that treats the robot as a particle in an artificial field where the goal exerts attraction and obstacles exert repulsion, with the robot descending the combined gradient. Introduced for real-time obstacle avoidance by Khatib in 1986, it is fast and naturally reactive but suffers from local minima, motivating variants such as navigation functions and harmonic potentials. Why it matters for physical AI: Potential fields remain a building block for reactive collision avoidance layers that wrap learned or planned motions, and their local-minima failure mode is a standing lesson in why purely reactive schemes need global guidance. ### Pouring Category: Manipulation URL: https://almond.bot/glossary/pouring Pouring is the manipulation task of transferring liquid or granular material between containers, requiring control of container tilt and trajectory in response to flow. It is a standard benchmark for dynamic, non-rigid manipulation because outcomes depend on material properties and fluid dynamics that are hard to model, and success requires perception of fill level and spillage, often from vision alone. Why it matters for physical AI: Household and industrial tasks are full of liquids and granular media that rigid-body assumptions cannot capture, making pouring a compact test of whether policies can handle dynamics beyond rigid objects. ### Power Density Category: Hardware URL: https://almond.bot/glossary/power-density Power density is the ratio of deliverable mechanical power to mass, typically in watts per kilogram, used to compare actuators and complete robots. Hydraulics historically led in raw power density, which enabled dynamic machines like early Atlas, while modern high-torque electric motors with low-ratio gearing now approach them with far better efficiency and controllability. System-level power density also depends on batteries, drives, and thermal limits. Why it matters for physical AI: Dynamic behaviors such as running, jumping, and heavy lifting are gated by actuator power density, and battery plus actuator power budgets fundamentally shape humanoid and quadruped design trade-offs. ### Pre-Training Category: Foundation Models URL: https://almond.bot/glossary/pre-training Pre-training is the initial training of a model on a large, broad dataset to acquire general representations before fine-tuning on a downstream task. In robotics this takes several forms: visual encoders pretrained on human video as in R3M and VC-1, cross-embodiment pre-training on pooled robot datasets such as Open X-Embodiment, and initializing vision-language-action models from web-pretrained vision-language backbones. Why it matters for physical AI: Robot data is scarce and expensive, so leveraging knowledge pretrained on web and cross-robot corpora is the central strategy by which robot foundation models escape the limits of per-task data collection. ### Precision Category: Hardware URL: https://almond.bot/glossary/precision Precision is the degree to which a robot reproduces the same motion or measurement under unchanged conditions, and it is closely tied to repeatability, whereas accuracy measures closeness to the commanded target. ISO 9283 defines the standard pose accuracy and repeatability tests for industrial manipulators. A robot can be precise but inaccurate, which is why kinematic calibration is used to convert repeatability into usable accuracy. Why it matters for physical AI: Whether a task needs micrometer precision or centimeter tolerance determines viable hardware, sensing, and control strategies, and conflating precision with accuracy is a classic source of failed automation specifications. ### Precision Agriculture Category: Applications URL: https://almond.bot/glossary/precision-agriculture Precision agriculture is the practice of managing crops and livestock at fine spatial resolution using GNSS guidance, remote sensing, and increasingly robotics, applying inputs such as water, fertilizer, and herbicide only where needed. Robotic embodiments include auto-steered tractors, drone-based crop monitoring, mechanical weeding robots, targeted micro-spraying, and harvesting robots for fruit and vegetables that combine detection with delicate manipulation. Why it matters for physical AI: Agriculture offers vast unstructured outdoor environments, severe labor shortages, and per-plant tasks that demand robust perception and gentle manipulation, making it a proving ground for field-deployable autonomy. ### Predictive Functional Control Category: Control URL: https://almond.bot/glossary/predictive-functional-control Predictive functional control (PFC) is a simplified form of model predictive control, developed by Jacques Richalet, that parameterizes the future control input as a weighted sum of simple basis functions and enforces reference tracking only at a few coincidence points on the prediction horizon. The resulting computation is light enough for fast industrial loops, which made PFC one of the earliest widely fielded predictive control techniques. Why it matters for physical AI: PFC illustrates a recurring deployment lesson: aggressive structural simplification can retain most of predictive control's benefit at a fraction of the compute, an approach echoed in today's real-time MPC for legged robots. ### Preference Learning Category: Robot Learning URL: https://almond.bot/glossary/preference-learning Preference learning is the inference of objectives from comparative judgments, typically a human indicating which of two trajectories better accomplishes a task, rather than from hand-specified rewards or demonstrations. Christiano et al. (2017) showed that deep reinforcement learning from such pairwise preferences can train complex behaviors from surprisingly few queries, establishing the reward-model recipe later scaled into RLHF for language models. Why it matters for physical AI: Many robot objectives, such as gentle handling or socially acceptable motion, are easier to recognize than to formalize, and preference queries give non-experts a practical channel for specifying them. ### Prehensile Manipulation Category: Manipulation URL: https://almond.bot/glossary/prehensile-manipulation Prehensile manipulation is manipulation performed while grasping, with the object held or caged by the hand so its motion is controlled through the grasp. It contrasts with nonprehensile manipulation, such as pushing, sliding, and tossing, where the object is moved without being held. Most industrial and learned manipulation is prehensile, since a firm grasp reduces the object's uncertainty to that of the end effector. Why it matters for physical AI: The prehensile versus nonprehensile distinction structures the manipulation design space: grasping buys certainty at the cost of grasp feasibility, while contact-rich alternatives extend what simple grippers can achieve. ### Pretrained Visual Representations Category: Foundation Models URL: https://almond.bot/glossary/pretrained-visual-representations Pretrained visual representations (PVRs) are vision encoders trained on large image or video corpora and reused, frozen or fine-tuned, as the perception backbone of robot policies. Prominent examples include R3M, trained on Ego4D egocentric video with time-contrastive and language-alignment objectives, MVP and VC-1 based on masked autoencoding, and CLIP-style encoders. Benchmarks such as CortexBench show their benefits vary considerably across tasks and evaluation conditions. Why it matters for physical AI: Reusing encoders pretrained on abundant human video sidesteps the scarcity of robot data for perception, though choosing and adapting the right representation for control remains an empirical, task-dependent decision. ### Primitive Actions Category: Robot Learning URL: https://almond.bot/glossary/primitive-actions Primitive actions are the low-level building-block behaviors from which longer robot activities are composed, ranging from atomic motor commands to parameterized skills such as reach, grasp, push, or screw. Formulations include dynamic movement primitives that encode motions as stable dynamical systems, options in hierarchical reinforcement learning, and skill libraries invoked by task planners or, increasingly, by language-model-based high-level policies. Why it matters for physical AI: Choosing the action abstraction is a core design decision for robot foundation models: composing proven primitives improves reliability and sample efficiency, while end-to-end low-level control preserves flexibility that fixed primitives lose. ### Privileged Information Category: Robot Learning URL: https://almond.bot/glossary/privileged-information Privileged information is ground-truth state available during training but not at deployment, such as exact object poses, contact forces, terrain friction, or full maps available inside a simulator. The teacher-student paradigm exploits it: a teacher policy trained with privileged access supervises a student restricted to deployable sensors, as in Learning by Cheating for driving and Lee et al.'s 2020 quadruped controller that inferred terrain from proprioceptive history. Why it matters for physical AI: Privileged training decouples what is hard to learn from what is hard to sense, and it underlies many of the most successful sim-to-real systems in locomotion and manipulation. ### Probabilistic Roadmap (PRM) Category: Navigation & SLAM URL: https://almond.bot/glossary/probabilistic-roadmap A probabilistic roadmap (PRM) is a sampling-based motion planner, introduced by Kavraki, Svestka, Latombe, and Overmars in 1996, that randomly samples collision-free configurations, connects nearby samples with feasible local paths to form a graph, and answers queries by graph search. As a multi-query method it amortizes roadmap construction across many planning problems, and the PRM* variant achieves asymptotic optimality. Why it matters for physical AI: Sampling-based planners like PRM made motion planning tractable in high-dimensional configuration spaces, and precomputed roadmaps still power fast repeated planning in static industrial workcells. ### Procedural Generation Category: Simulation URL: https://almond.bot/glossary/procedural-generation Procedural generation is the algorithmic creation of simulation content, including scene layouts, object arrangements, textures, and task variations, rather than hand-authoring each environment. Systems such as ProcTHOR generate thousands of household layouts, and RoboCasa combines procedural kitchens with generative-AI-created assets. It extends domain randomization from perturbing parameters to synthesizing structurally novel environments at scale. Why it matters for physical AI: Diversity of training environments is a primary driver of policy generalization, and procedural pipelines are the only practical way to produce the millions of distinct scenes large-scale simulation training consumes. ### Projective Transformation Category: Math & Kinematics URL: https://almond.bot/glossary/projective-transformation A projective transformation, or homography, is the most general invertible linear map on homogeneous coordinates, representable in the plane as a 3x3 matrix with eight degrees of freedom that preserves straight lines but not parallelism or ratios. Perspective imaging is projective, and homographies exactly relate two views of a planar scene or two views from a purely rotating camera. Why it matters for physical AI: Homographies underpin practical robot vision: fiducial marker pose estimation, ground-plane mapping for mobile robots, image rectification, and planar visual servoing all reduce to estimating and applying projective transforms. ### Proprioception Category: Perception URL: https://almond.bot/glossary/proprioception Proprioception is a robot's sensing of its own internal state, including joint positions and velocities from encoders, actuator torques or motor currents, and base orientation and acceleration from an IMU. It contrasts with exteroception, the sensing of the external world. Proprioceptive signals are high-rate, low-latency, and reliable in conditions that defeat cameras, and histories of them implicitly encode contact events and terrain properties. Why it matters for physical AI: Blind quadruped locomotion over rough terrain from proprioceptive history alone showed how much environmental information internal sensing carries, and virtually every learned policy consumes proprioception alongside vision. ### Prosthetic Limb Category: Applications URL: https://almond.bot/glossary/prosthetic-limb A prosthetic limb is an artificial replacement for a missing arm or leg, with robotic versions adding powered joints, embedded sensing, and microprocessor control. Myoelectric prostheses decode electromyographic signals from residual muscles to command grasp and joint motion, while research systems pursue targeted muscle reinnervation, implanted neural interfaces, and sensory feedback through electrotactile or neural stimulation to restore touch. Why it matters for physical AI: Prosthetics compress core physical AI challenges, including intent decoding, compliant actuation, and closed-loop haptic feedback, into a wearable, safety-critical form factor operating continuously alongside a human body. ### Proximal Policy Optimization (PPO) Category: Robot Learning URL: https://almond.bot/glossary/proximal-policy-optimization Proximal policy optimization (PPO) is a policy-gradient reinforcement learning algorithm introduced by Schulman et al. in 2017 that stabilizes training by clipping the probability ratio between the updated and previous policy, approximating a trust-region constraint with a simple first-order objective. Its robustness to hyperparameters and easy parallelization made it the default algorithm for large-scale simulated robot learning. Why it matters for physical AI: PPO paired with massively parallel GPU simulation is the workhorse recipe behind most modern sim-to-real locomotion results, from quadruped parkour to humanoid whole-body control. ### Proximity Sensor Category: Hardware URL: https://almond.bot/glossary/proximity-sensor A proximity sensor detects the presence or distance of nearby objects without contact, using inductive, capacitive, ultrasonic, infrared, or time-of-flight principles. Inductive sensors detect metal at short range and dominate industrial end-of-travel detection, while optical and capacitive variants extend to nonmetallic targets. On robots they serve limit sensing, gripper part-presence checks, collision-avoidance skins, and pre-touch sensing just before grasp contact. Why it matters for physical AI: Cheap, fast, non-contact presence sensing fills the gap between vision and touch, and proximity-sensing robot skins are a practical route to safe high-speed operation near people. ### Pseudoinverse Category: Math & Kinematics URL: https://almond.bot/glossary/pseudoinverse The pseudoinverse, formally the Moore-Penrose generalized inverse, extends matrix inversion to non-square and rank-deficient matrices, yielding the least-squares solution of minimum norm. In robotics its central use is the Jacobian pseudoinverse, which maps desired end-effector velocities to joint velocities for redundant manipulators, with damped least-squares variants trading tracking accuracy for bounded joint speeds near singularities. Why it matters for physical AI: Jacobian-pseudoinverse mappings run inside virtually every Cartesian controller and teleoperation pipeline, and their singularity behavior explains characteristic failures such as joint-velocity blowups near workspace boundaries. ### Push Manipulation Category: Manipulation URL: https://almond.bot/glossary/push-manipulation Push manipulation is the nonprehensile repositioning of objects by pushing them across a support surface rather than grasping them. Mason's 1986 analysis of pushing mechanics founded the field, characterizing how objects rotate and translate under frictional point and edge contacts. Pushing serves to singulate cluttered objects, move items too large or flat to grasp, and reposition objects into graspable poses. Why it matters for physical AI: Pushing is the simplest contact-rich skill exhibiting the uncertainty and hybrid dynamics that make manipulation hard, which is why planar pushing remains a favored benchmark for learned dynamics models and policies. ### Push Recovery Category: Locomotion URL: https://almond.bot/glossary/push-recovery Push recovery is a legged robot's ability to reject external perturbations and avoid falling, using a hierarchy of strategies: ankle torque for small pushes, hip and angular-momentum strategies for moderate ones, and reactive stepping for large disturbances. Capture point theory formalizes where a robot must step to come to rest, while modern learned controllers acquire recovery behaviors implicitly through randomized perturbations in simulation. Why it matters for physical AI: Robustness to bumps, slips, and contact is a prerequisite for deploying humanoids around people, and standardized push tests have become a de facto benchmark for locomotion controller quality. ### Push-T Benchmark Category: Data & Benchmarks URL: https://almond.bot/glossary/push-t-benchmark Push-T is a planar manipulation benchmark in which a policy must push a T-shaped block to a target pose using a circular pusher, adapted from the Implicit Behavioral Cloning task suite and popularized by the Diffusion Policy paper of Chi et al. (2023). The task stresses multimodal solution strategies, contact-rich dynamics, and precise long-horizon corrections, and it is evaluated by final pose overlap with the goal. Why it matters for physical AI: Push-T became the standard quick test of whether a policy class can represent multimodal action distributions, and results on it drove the adoption of diffusion-based policies for manipulation. ### PyBullet Category: Simulation URL: https://almond.bot/glossary/pybullet PyBullet is the Python interface to the Bullet physics engine, developed by Erwin Coumans and Yunfei Bai, providing rigid-body simulation, collision checking, forward and inverse kinematics and dynamics, and URDF loading for robotics research. Free, lightweight, and easy to install, it powered a generation of robot learning benchmarks and sim-to-real studies before GPU-parallel simulators arrived, and it remains popular for prototyping. Why it matters for physical AI: Accessible simulation lowered the entry barrier for robot learning research, and many reference implementations and benchmarks still target PyBullet, keeping it relevant for reproducing and extending prior work. ### PyTorch Category: Software & Middleware URL: https://almond.bot/glossary/pytorch PyTorch is an open-source deep learning framework, originated at Meta AI in 2016, that builds computation graphs dynamically and provides automatic differentiation, GPU acceleration, and a large ecosystem of pretrained models. It dominates robot learning research: most vision-language-action models, diffusion policies, and reinforcement learning libraries, including projects like LeRobot, are implemented in it, with TorchScript and torch.compile supporting deployment. Why it matters for physical AI: Framework choice shapes the research-to-robot pipeline, and PyTorch's dominance means robot learning inherits its tooling for distributed training, model exchange, and increasingly for optimized on-robot inference. ## Glossary: Q ### Q-Function (Action-Value Function) Category: Robot Learning URL: https://almond.bot/glossary/q-function A Q-function (action-value function) is the expected cumulative discounted reward obtained by taking a given action in a given state and following a policy thereafter. It satisfies the Bellman equation, and the optimal Q-function directly induces an optimal policy through greedy action selection. Q-functions are the central object of value-based reinforcement learning, estimated with neural networks in deep RL methods such as DQN and SAC. Why it matters for physical AI: Value estimates let robots rank candidate actions without exhaustive physical trials, and learned Q-functions underpin offline RL, grasp quality scoring, and policy improvement on real hardware. ### Q-Learning Category: Robot Learning URL: https://almond.bot/glossary/q-learning Q-learning is an off-policy reinforcement learning algorithm, introduced by Watkins in 1989, that learns the optimal action-value function through temporal-difference updates toward the observed reward plus the discounted maximum Q-value of the next state. In the tabular setting it converges to the optimal policy regardless of the exploration policy used. Deep Q-Networks (DQN) scaled the idea to high-dimensional inputs in 2015. Why it matters for physical AI: Off-policy learning means data from any source, including old policies and human demonstrations, can improve the value function, a property robotics depends on given the cost of real-world experience. ### Quadratic Program (QP) Category: Math & Kinematics URL: https://almond.bot/glossary/quadratic-program A quadratic program (QP) is an optimization problem with a quadratic objective and linear constraints, convex and efficiently solvable when the objective's Hessian is positive semidefinite. Robotics runs QPs at control rates for whole-body control, inverse kinematics, contact force distribution, and control-barrier-function safety filters, using solvers such as OSQP, qpOASES, and ProxQP. Why it matters for physical AI: The QP is the workhorse that turns high-level commands, including learned policy outputs, into constraint-satisfying joint torques hundreds of times per second on legged and manipulator systems. ### Quadruped Category: Locomotion URL: https://almond.bot/glossary/quadruped A quadruped is a four-legged robot, a form factor that balances stability, agility, and mechanical simplicity relative to bipeds, with prominent examples including Boston Dynamics Spot, ANYbotics ANYmal, and the Unitree and MIT Mini Cheetah lines. Quadrupeds became the proving ground for learned locomotion, where reinforcement learning policies trained in simulation now routinely outperform hand-designed controllers on rough terrain. Why it matters for physical AI: Quadrupeds delivered the first commercially deployed learned locomotion at scale, and lessons from their sim-to-real pipelines now inform humanoid and mobile manipulation training. ### Quality Control Robot Category: Applications URL: https://almond.bot/glossary/quality-control-robot A quality control robot is a robot that automates inspection and testing in production, using cameras, 3D scanners, or specialized probes to detect defects, verify dimensions, and confirm assembly correctness. Deployments range from fixed vision stations to arms that reposition sensors around complex parts, increasingly powered by learned defect-detection models that adapt to new part types from few examples. Why it matters for physical AI: Inspection is a high-value early market for perception-heavy robotics: tasks are repetitive and measurable, and anomaly-detection advances transfer directly from mainstream computer vision. ### Quality-Diversity Optimization Category: Robot Learning URL: https://almond.bot/glossary/quality-diversity-optimization Quality-diversity optimization is a family of evolutionary methods that seeks a collection of solutions that are both high-performing and behaviorally diverse, rather than a single optimum. MAP-Elites, the best-known algorithm, maintains an archive of elite solutions across a grid of behavior descriptors. Cully et al.'s 2015 Nature paper used such archives to let a damaged hexapod rapidly find compensating gaits. Why it matters for physical AI: Behavior repertoires give robots fallback strategies when hardware degrades or environments shift, and diverse solution archives seed adaptation far faster than reoptimizing from scratch. ### Quantization (Model Quantization) Category: Foundation Models URL: https://almond.bot/glossary/quantization Quantization (model quantization) is the conversion of a neural network's weights and activations from high-precision floating point to lower-precision formats such as INT8, FP8, or 4-bit representations, shrinking memory footprint and accelerating inference. Post-training quantization converts a finished model directly, while quantization-aware training simulates low precision during training to preserve accuracy. Why it matters for physical AI: Billion-parameter vision-language-action models only fit real-time control budgets on embedded hardware after aggressive quantization, making it a standard step in deploying robot foundation models onboard. ### Quasi-Direct Drive (QDD) Category: Hardware URL: https://almond.bot/glossary/quasi-direct-drive A quasi-direct drive (QDD) is an actuator that pairs a high-torque electric motor with a low gear ratio, typically around 6:1 to 10:1, preserving much of a direct drive's backdrivability and transparency while multiplying torque. Popularized by the MIT Cheetah robots, QDD actuators enable proprioceptive force sensing, estimating contact forces from motor current without dedicated sensors, and tolerate impacts that damage highly geared drives. Why it matters for physical AI: Cheap, impact-robust force sensitivity is exactly what dynamic legged robots and contact-rich manipulators need, which is why QDD actuation spread from research quadrupeds to commercial platforms. ### Quasi-Static Manipulation Category: Manipulation URL: https://almond.bot/glossary/quasi-static-manipulation Quasi-static manipulation is manipulation performed slowly enough that inertial forces are negligible, so the system passes through a sequence of force-balanced equilibrium states. The assumption reduces the physics to a balance of contact, gravity, and applied forces, making analysis and planning of tasks like pushing, grasping, and part reorientation far more tractable. It contrasts with dynamic manipulation such as throwing or catching. Why it matters for physical AI: Most practical manipulation runs in the quasi-static regime, and the assumption underlies many grasp analyses, contact planners, and stability checks used in deployed systems. ### Quaternion Category: Math & Kinematics URL: https://almond.bot/glossary/quaternion A quaternion is a four-dimensional number whose unit elements represent 3D rotations compactly and without the gimbal-lock singularities of Euler angles. Unit quaternions double-cover the rotation group SO(3), with q and -q encoding the same rotation, and support smooth interpolation via slerp. They are the standard orientation representation in robot state estimation, control, and simulation. Why it matters for physical AI: Orientation representation choices ripple through estimators, controllers, and learned models; quaternions' compactness and singularity-freedom make them the default, though the double cover demands care in losses and filters. ### Quest 3 Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/quest-3-teleoperation Quest 3 teleoperation is the use of Meta's consumer Quest 3 mixed-reality headset as a robot teleoperation interface, streaming the operator's head pose, controller poses, or camera-based hand tracking as command signals while displaying the robot's camera feed in the headset. Open-source stacks built on it, such as Open-TeleVision and related systems, made immersive bimanual teleoperation available at consumer hardware prices. Why it matters for physical AI: Demonstration data is the fuel of imitation-learned policies, and sub-thousand-dollar VR teleoperation rigs dramatically widened who can collect high-quality bimanual manipulation data. ## Glossary: R ### R3M Category: Foundation Models URL: https://almond.bot/glossary/r3m R3M (Reusable Representations for Robot Manipulation) is a pretrained visual encoder introduced by Nair et al. in 2022 that trains a ResNet on the Ego4D egocentric human video dataset using time-contrastive learning and video-language alignment objectives. Used as a frozen perception backbone, R3M features improved downstream imitation-learning success over training from scratch or ImageNet features across simulated and real manipulation tasks. Why it matters for physical AI: R3M was among the first demonstrations that representations learned from human video transfer to robot control, helping establish video pre-training as a route around scarce robot data. ### RANSAC Category: Perception URL: https://almond.bot/glossary/ransac RANSAC (random sample consensus) is a robust estimation algorithm introduced by Fischler and Bolles in 1981 that fits models to outlier-contaminated data by repeatedly sampling minimal subsets, hypothesizing a model from each, and keeping the hypothesis with the largest inlier consensus. In robotics it underpins plane extraction from point clouds, feature-based pose and homography estimation, and geometric verification of loop-closure candidates. Why it matters for physical AI: Real sensor data is full of gross outliers from mismatches, reflections, and dynamic objects, and RANSAC-style hypothesize-and-verify remains the default defense wherever geometric models are fit to it. ### Rapidly-Exploring Random Tree (RRT) Category: Navigation & SLAM URL: https://almond.bot/glossary/rapidly-exploring-random-tree A rapidly-exploring random tree (RRT) is a sampling-based motion planner, introduced by LaValle in 1998, that incrementally grows a tree from the start configuration by sampling random configurations and extending the nearest tree node toward each sample. Its Voronoi-biased growth explores high-dimensional spaces quickly; RRT-Connect grows bidirectional trees for speed, and RRT* (Karaman and Frazzoli, 2011) adds rewiring for asymptotic optimality. Why it matters for physical AI: RRT variants remain the default single-query planners for manipulator motion generation, and they routinely serve as the classical baseline and fallback layer around learned motion policies. ### RDT-1B Category: Foundation Models URL: https://almond.bot/glossary/rdt-1b RDT-1B (Robotics Diffusion Transformer) is a roughly 1.2-billion-parameter diffusion-based foundation model for bimanual manipulation, released by Tsinghua University researchers in 2024. It is pretrained on large multi-robot datasets using a unified action space that accommodates heterogeneous embodiments, then fine-tuned on a self-collected dataset of over 6,000 bimanual episodes on an ALOHA platform, demonstrating language-conditioned dexterous tasks and few-shot adaptation. Why it matters for physical AI: RDT-1B showed that diffusion policies scale to foundation-model size and that unified action representations enable cross-embodiment pre-training, both influential design choices for bimanual robot learning. ### Reach Category: Hardware URL: https://almond.bot/glossary/reach Reach is the maximum distance a robot's wrist or tool flange can extend from its base axis, defining the outer boundary of its workspace. Manufacturers quote horizontal and vertical reach, and the usable dexterous workspace, where full orientation control is available, is smaller than the reachable envelope. Reach trades off against stiffness, payload, and precision, since longer links amplify deflection and inertia. Why it matters for physical AI: Reach determines workcell layout, part presentation, and whether a fixed arm can service a task at all, and it interacts with mobile bases that extend effective reach at the cost of localization uncertainty. ### Reachability Category: Math & Kinematics URL: https://almond.bot/glossary/reachability Reachability is the property of whether a robot can attain a given pose or state, and reachability analysis computes the set of attainable end-effector poses or system states. In manipulation, precomputed reachability or capability maps score how well regions of the workspace can be reached across orientations, guiding grasp selection and mobile-base placement. In safety-oriented control, reachable-set computation bounds where a system can end up under bounded inputs and disturbances. Why it matters for physical AI: Reachability reasoning connects kinematics to decision-making: it tells a mobile manipulator where to stand before acting, and it gives formal safety methods the machinery to certify collision and constraint avoidance. ### Reactive Policy Category: Robot Learning URL: https://almond.bot/glossary/reactive-policy A reactive policy is a controller that maps the current observation, or a short history, directly to an action without explicit planning, search, or internal simulation at decision time. Most end-to-end learned policies, from behavior-cloned manipulation networks to RL locomotion controllers, are reactive, achieving high control rates and robustness to disturbances while relying on training-time experience rather than deliberation to handle long-horizon structure. Why it matters for physical AI: Fast closed-loop reaction is what lets learned controllers correct errors and reject perturbations in real time, but the deliberation-free design also explains their struggles with novel long-horizon tasks, motivating hybrid planner-policy architectures. ### Real-Time Control Category: Control URL: https://almond.bot/glossary/real-time-control Real-time control is control computation subject to hard timing deadlines, where a late result is a wrong result. Robot servo loops typically run at 1 to 10 kHz, whole-body and impedance controllers at hundreds of hertz, and missed deadlines or jitter degrade stability margins. Achieving determinism requires real-time operating systems, deterministic fieldbuses such as EtherCAT, and careful isolation from non-deterministic workloads like vision or neural network inference. Why it matters for physical AI: Learned components must coexist with hard real-time loops, so architectures that pair high-rate deterministic control with slower policy inference, connected through action chunking or setpoint streaming, are central to deploying neural policies on physical robots. ### Real-Time Operating System (RTOS) Category: Software & Middleware URL: https://almond.bot/glossary/real-time-operating-system A real-time operating system (RTOS) is an operating system that provides deterministic scheduling guarantees, ensuring high-priority tasks meet deadlines with bounded latency through preemptive priority scheduling, priority-inheritance locking, and minimal interrupt jitter. Examples include VxWorks, QNX, and FreeRTOS, while Linux with the PREEMPT_RT patches achieves soft-to-firm real-time performance and is widely used in robot controllers. Why it matters for physical AI: Deterministic execution is the substrate beneath every stable high-rate control loop, and choosing between microcontroller RTOSes and real-time Linux shapes a robot's compute architecture and its integration of learned components. ### Real-to-Sim Transfer Category: Simulation URL: https://almond.bot/glossary/real-to-sim-transfer Real-to-sim transfer is the construction or calibration of simulation environments from real-world data, reconstructing scene geometry and appearance with photogrammetry, NeRFs, or Gaussian splatting, and identifying dynamics parameters such as mass, friction, and actuator response through system identification. The resulting digital twins support policy training and evaluation matched to a specific deployment site, closing the loop in real-to-sim-to-real pipelines. Why it matters for physical AI: Building simulators from real scenes attacks the sim-to-real gap at its source, and scene-specific digital twins enable safe policy evaluation and targeted fine-tuning before touching the physical system. ### Rectification Category: Perception URL: https://almond.bot/glossary/rectification Rectification is the warping of a stereo image pair so that corresponding epipolar lines become horizontal and row-aligned, reducing stereo correspondence search from two dimensions to one. Computed from calibrated camera intrinsics, extrinsics, and lens distortion models, it is the standard preprocessing step before block matching or learned stereo disparity estimation. The term also covers single-image undistortion to an ideal pinhole model. Why it matters for physical AI: Nearly every depth camera and stereo pipeline on a robot rectifies images before matching, and calibration drift that invalidates rectification is a common silent cause of degraded depth and downstream grasp failures. ### Recurrent Neural Network (RNN) Category: Foundation Models URL: https://almond.bot/glossary/recurrent-neural-network A recurrent neural network (RNN) is a neural architecture with feedback connections that maintain a hidden state across time steps, allowing it to process sequences of arbitrary length with constant memory. Gated variants such as LSTM and GRU mitigate vanishing gradients. Although transformers now dominate sequence modeling, recurrent policies remain attractive on robots for their constant per-step compute and their ability to integrate observation histories under partial observability. Why it matters for physical AI: Memory lets a policy infer unobserved quantities, such as terrain friction or object mass, from interaction history, and recurrent state-estimation modules still feature in deployed locomotion controllers where inference budgets are tight. ### Reduced-Order Model Category: Control URL: https://almond.bot/glossary/reduced-order-model A reduced-order model is a simplified dynamics model that captures the dominant behavior of a complex system with far fewer states, trading fidelity for tractability. Legged locomotion classics include the linear inverted pendulum for bipedal walking, the spring-loaded inverted pendulum for running, and single-rigid-body models for quadrupeds, each enabling real-time model predictive control that full-order dynamics would make computationally prohibitive. Why it matters for physical AI: Real-time planning on legged robots exists largely because reduced-order models make the optimization tractable, and template-and-anchor hierarchies remain the backbone of model-based humanoid control. ### Redundancy Resolution Category: Math & Kinematics URL: https://almond.bot/glossary/redundancy-resolution Redundancy resolution is the selection of a specific joint motion when a manipulator has more degrees of freedom than its task requires, leaving infinitely many inverse kinematics solutions. Standard techniques project secondary objectives, such as joint-limit avoidance, singularity avoidance, obstacle clearance, or posture preference, into the null space of the task Jacobian so they proceed without disturbing the primary end-effector motion. Why it matters for physical AI: Seven-axis arms and humanoid torsos are now common precisely because redundancy buys dexterity and obstacle avoidance, but exploiting it requires principled resolution, whether by null-space control or learned whole-body policies. ### Regrasping Category: Manipulation URL: https://almond.bot/glossary/regrasping Regrasping is the deliberate change of a robot's grasp on an object, either by placing the object down and picking it up differently, transferring it between two hands or arms, or manipulating it within the hand. It is required when the initial feasible grasp is incompatible with the downstream task, such as inserting a part that was grasped over its mating surface, and it is a classic planning problem combining grasp and placement reasoning. Why it matters for physical AI: Regrasp competence separates single-shot pick-and-place from genuine task versatility, and dual-arm handovers and in-hand regrasps are increasingly prominent skills in bimanual learning benchmarks. ### Rehabilitation Robot Category: Applications URL: https://almond.bot/glossary/rehabilitation-robot A rehabilitation robot is a robotic system that assists recovery of motor function after stroke, spinal cord injury, or other neurological impairment, delivering high-dosage repetitive therapy. Categories include treadmill-based gait exoskeletons such as the Lokomat, end-effector arm trainers descended from MIT-Manus, and wearable powered orthoses. Assist-as-needed control adapts robot support to patient effort, since active participation drives neuroplasticity and recovery. Why it matters for physical AI: Rehabilitation demands compliant, human-safe interaction control with adaptive assistance in continuous physical contact with a person, concentrating the hardest problems of physical human-robot interaction into one application. ### Reinforcement Learning for Robotics Category: Robot Learning URL: https://almond.bot/glossary/reinforcement-learning-for-robotics Reinforcement learning for robotics is the application of trial-and-error learning to robot control, training policies to maximize cumulative reward through interaction. Because real-robot experience is slow, costly, and unsafe, the dominant recipe trains in massively parallel simulation with domain randomization and transfers to hardware. Landmark results include OpenAI's in-hand Rubik's cube manipulation, learned quadruped locomotion over wild terrain, and champion-level autonomous drone racing. Why it matters for physical AI: Reinforcement learning produces behaviors no one can demonstrate or hand-program, particularly dynamic whole-body skills, making it the complement to imitation learning in the modern robot learning toolbox. ### Reinforcement Learning from Human Feedback (RLHF) Category: Foundation Models URL: https://almond.bot/glossary/reinforcement-learning-from-human-feedback Reinforcement learning from human feedback (RLHF) is a training paradigm that fits a reward model to human preference comparisons between behavior samples, then optimizes a policy against that learned reward with reinforcement learning. Building on the preference-learning work of Christiano et al. (2017), RLHF became the standard alignment stage for large language models and is being adapted to align robot behavior with hard-to-specify human expectations. Why it matters for physical AI: Reward specification is a chronic failure point in robot RL, and preference-based reward models offer a scalable way to encode judgments like gentleness or appropriateness that resist manual formalization. ### Remote Presence Robot Category: Teleoperation URL: https://almond.bot/glossary/remote-presence-robot A remote presence robot, or telepresence robot, is a mobile platform carrying a screen, camera, microphone, and speakers that lets a remote operator see, hear, converse, and move through a distant environment. Deployed in hospitals for specialist consultations, in offices for remote workers, and in homes for family connection, most are drive-only platforms, distinguishing them from teleoperated manipulators that also act on the world. Why it matters for physical AI: Telepresence is a commercially proven slice of teleoperation that exposes the latency, situational awareness, and social acceptance challenges any remotely supervised robot fleet must solve. ### Repeatability Category: Hardware URL: https://almond.bot/glossary/repeatability Repeatability is the closeness of agreement between repeated robot poses attained under identical commands and conditions, standardized by ISO 9283 as pose repeatability. Industrial arms typically achieve plus or minus 0.01 to 0.1 mm. It is distinct from accuracy, the error relative to an absolutely commanded target, which is usually far worse until kinematic calibration corrects systematic model errors. Why it matters for physical AI: Teach-and-repeat automation runs on repeatability, but perception-driven and offline-programmed systems need absolute accuracy, so the distinction determines whether a deployment requires calibration, vision correction, or neither. ### Replay Buffer Category: Robot Learning URL: https://almond.bot/glossary/replay-buffer A replay buffer is a memory that stores past interaction transitions, typically tuples of state, action, reward, and next state, from which off-policy reinforcement learning algorithms sample minibatches for updates. Introduced as experience replay and popularized by DQN, it decorrelates sequential samples and reuses experience many times, with prioritized replay weighting samples by learning progress. Its logic extends naturally to offline RL on fixed datasets. Why it matters for physical AI: Sample reuse is precious when experience comes from physical hardware, and replay-based off-policy methods are the standard choice for the modest data budgets of real-robot reinforcement learning. ### Representation Learning Category: Robot Learning URL: https://almond.bot/glossary/representation-learning Representation learning is the acquisition of feature embeddings from raw sensor data that expose task-relevant structure, typically without manual labels, through objectives such as contrastive learning, masked reconstruction, temporal prediction, or language alignment. In robotics it spans visual encoders pretrained on human video, state representations learned from interaction, and latent dynamics models, all aiming to make downstream policy learning easier and more sample-efficient. Why it matters for physical AI: Good representations are the leverage point of data-limited robot learning: a policy trained on a few hundred demonstrations succeeds or fails largely on what its pretrained encoder already understands. ### Reset-Free Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/reset-free-reinforcement-learning Reset-free reinforcement learning is the training of RL agents in the real world without manual environment resets between episodes, which otherwise demand constant human intervention. Approaches include jointly learning a reset or recovery policy that returns the environment to viable start states, cyclic multi-task curricula in which each task resets another, and formulating learning as one continual autonomous process rather than discrete episodes. Why it matters for physical AI: Human-supervised resets are the hidden labor cost that keeps RL off real robots at scale, so reset-free autonomy is a prerequisite for robots that improve unattended from their own experience. ### Residual Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/residual-reinforcement-learning Residual reinforcement learning trains a policy to output corrective actions that are added to a base controller, such as a hand-designed feedback law or model-based planner, rather than learning control from scratch. Introduced concurrently by Johannink et al. and Silver et al. around 2018, the decomposition lets the base controller handle nominal behavior and safety while the learned residual absorbs unmodeled effects, excelling in contact-rich tasks like insertion. Why it matters for physical AI: Starting from a competent controller slashes exploration cost and risk on physical hardware, making residual learning one of the most deployment-friendly ways to add learning to existing industrial systems. ### ResNet Category: Foundation Models URL: https://almond.bot/glossary/resnet ResNet is a convolutional neural network architecture introduced by He et al. in 2015 whose residual skip connections let gradients flow through very deep networks, enabling models of 50 to 152 layers and winning the 2015 ImageNet challenge. In robotics, compact variants such as ResNet-18 and ResNet-50 remain standard visual encoders for policies, including R3M's backbone and the image encoders of many diffusion-policy implementations. Why it matters for physical AI: Convolutional backbones still offer strong accuracy per FLOP for onboard inference, and much of the pretrained-representation ecosystem robot learning depends on was built on ResNet features. ### Resolved Rate Control Category: Control URL: https://almond.bot/glossary/resolved-rate-control Resolved rate control is a Cartesian velocity control scheme, introduced by Whitney in 1969, that maps a desired end-effector twist to joint velocities through the inverse or pseudoinverse of the manipulator Jacobian at every control cycle. It underlies Cartesian jogging, visual servoing, and teleoperation rate control, with damped least-squares and null-space extensions handling singularities and redundancy. Why it matters for physical AI: Streaming end-effector velocity commands is the interface through which most teleoperation rigs and many learned policies drive arms, and resolved rate control is the machinery that makes that interface work. ### Resolver Category: Hardware URL: https://almond.bot/glossary/resolver A resolver is an electromagnetic rotary transducer that measures absolute shaft angle through the varying transformer coupling between a rotor winding and two stator windings in quadrature, whose outputs encode the sine and cosine of the shaft angle. Being brushless, contactless, and free of optics and onboard electronics, resolvers tolerate temperature extremes, vibration, and radiation that disable optical encoders, at the cost of resolution and signal-conditioning complexity. Why it matters for physical AI: Reliable joint sensing in harsh environments, from foundries to space, still favors resolvers, and their ruggedness illustrates the sensing trade-offs behind every actuator design choice. ### Reward Engineering Category: Robot Learning URL: https://almond.bot/glossary/reward-engineering Reward engineering is the design and iterative tuning of reward functions so that reinforcement learning produces the intended behavior, combining task success terms with shaping bonuses, energy and smoothness penalties, and constraint costs. It is notoriously error-prone: agents exploit loopholes in misspecified rewards, a failure mode known as reward hacking, and locomotion rewards in particular often require dozens of carefully weighted terms. Why it matters for physical AI: Reward design consumes a large share of practitioner effort in applied robot RL, and its brittleness motivates alternatives such as demonstrations, preference learning, and foundation-model-generated rewards. ### Reward Function Category: Robot Learning URL: https://almond.bot/glossary/reward-function A reward function is the scalar signal that defines a reinforcement learning objective, mapping each state and action, or resulting transition, to a numeric reward whose cumulative discounted sum the agent maximizes. It is the formal specification of the task within a Markov decision process. Sparse rewards, given only at success, are easy to specify but hard to explore under, while dense rewards guide learning at the risk of biasing the solution. Why it matters for physical AI: Everything an RL-trained robot does is downstream of its reward function, so the gap between the reward as written and the behavior actually desired is a first-order safety and reliability concern. ### Reward Model Category: Robot Learning URL: https://almond.bot/glossary/reward-model A reward model is a learned function that predicts reward or task success from observations, replacing or augmenting hand-specified reward signals. Reward models are trained from human preference comparisons, as in RLHF, from success-labeled examples, or increasingly by prompting vision-language models to judge task completion from images. On robots they also serve as autonomous success detectors that let systems label their own experience. Why it matters for physical AI: Scalable robot learning needs supervision that does not require a human watching every trial, and learned reward and success models are the main candidates for closing that autonomous feedback loop. ### Reward Shaping Category: Robot Learning URL: https://almond.bot/glossary/reward-shaping Reward shaping is the addition of auxiliary reward terms to a sparse task objective to guide exploration, such as rewarding distance reduction toward a goal or intermediate subtask completion. Ng, Harada, and Russell (1999) proved that potential-based shaping, where the bonus is a difference of a potential function over states, preserves the optimal policy, whereas arbitrary shaping can change what behavior is optimal and invite reward hacking. Why it matters for physical AI: Shaping is often what makes sparse-reward manipulation and locomotion tasks learnable at all, but the potential-based result is a standing warning that convenient bonuses can quietly redefine the task. ### RGB-D Camera Category: Hardware URL: https://almond.bot/glossary/rgb-d-camera An RGB-D camera is a sensor that outputs registered color and per-pixel depth images, using structured light as in the original Kinect, time-of-flight as in the Azure Kinect, or active infrared stereo as in the Intel RealSense D400 series. Depth range typically spans tens of centimeters to several meters, with characteristic failures on dark, reflective, or transparent surfaces and in direct sunlight. Why it matters for physical AI: Cheap registered depth made 3D manipulation perception ubiquitous, and RGB-D streams remain the standard input for grasp synthesis, obstacle avoidance, and the demonstration datasets behind most manipulation policies. ### Riemannian Geometry Category: Math & Kinematics URL: https://almond.bot/glossary/riemannian-geometry Riemannian geometry is the mathematics of smooth manifolds equipped with metrics that define distance, geodesics, and curvature. Robot orientation and pose spaces, such as the rotation group SO(3) and rigid-motion group SE(3), are curved manifolds, so interpolation, averaging, filtering, and optimization over them must respect geodesic structure rather than treat coordinates as Euclidean. Riemannian motion policies extend the framework to compose reactive controllers across task spaces. Why it matters for physical AI: Orientation errors, pose distributions, and trajectory optimization all live on curved manifolds, and geometry-aware formulations avoid the singularities and distortions that plague naive Euclidean treatments of rotation. ### Rigid Body Dynamics Category: Math & Kinematics URL: https://almond.bot/glossary/rigid-body-dynamics Rigid body dynamics is the study of forces and motion for systems of interconnected rigid links, expressed for manipulators by the equation M(q)q̈ + C(q, q̇)q̇ + g(q) = τ relating joint torques to accelerations. Efficient recursive algorithms, including the recursive Newton-Euler and Featherstone's articulated-body algorithm, evaluate inverse and forward dynamics in linear time and power both physics engines and model-based controllers, with libraries such as Pinocchio providing optimized implementations. Why it matters for physical AI: Dynamics models are the shared foundation of simulation, model predictive control, and system identification, and their fidelity limits everything built above them, from sim-to-real transfer to whole-body torque control. ### Risk Assessment Category: Safety & Standards URL: https://almond.bot/glossary/risk-assessment Risk assessment is the systematic identification, estimation, and evaluation of hazards posed by a robot system, followed by risk reduction measures, as required by machinery safety standards such as ISO 12100 and the industrial robot standards ISO 10218-1 and 10218-2. For collaborative applications, ISO/TS 15066 adds biomechanical force and pressure limits on human contact. The assessment determines required safeguarding, from fencing to speed and separation monitoring. Why it matters for physical AI: No robot ships into workplaces without a documented risk assessment, and the process increasingly must grapple with learned behaviors whose responses cannot be exhaustively enumerated in advance. ### Risk-Aware Reinforcement Learning Category: Robot Learning URL: https://almond.bot/glossary/risk-aware-reinforcement-learning Risk-aware reinforcement learning optimizes objectives sensitive to the distribution of returns rather than the expectation alone, using criteria such as conditional value-at-risk, worst-case formulations, or chance constraints on failure probability. Related machinery includes constrained MDPs with cost budgets, distributional RL that models full return distributions, and safety critics that veto high-risk actions during exploration and deployment. Why it matters for physical AI: A policy that is excellent on average but occasionally drops the object or falls is unacceptable on hardware around people, so tail-risk objectives speak directly to deployment readiness. ### RLBench Category: Data & Benchmarks URL: https://almond.bot/glossary/rlbench RLBench is a large-scale benchmark and learning environment for robot manipulation, introduced by James et al. in 2020, providing 100 language-described tasks of graded difficulty executed by a simulated Franka Panda arm in CoppeliaSim. Each task supplies procedurally varied scenes and an unlimited stream of scripted expert demonstrations, supporting imitation learning, reinforcement learning, and few-shot evaluation, and it became the standard testbed for language-conditioned agents such as PerAct. Why it matters for physical AI: Shared simulated benchmarks with demonstrations let manipulation methods be compared on equal footing, and RLBench's language-conditioned multi-task design anticipated how robot foundation models are now evaluated. ### RLDS Category: Data & Benchmarks URL: https://almond.bot/glossary/rlds RLDS (Reinforcement Learning Datasets) is a dataset format and toolchain from Google for recording, storing, and consuming episodic decision-making data without loss of information, built on TensorFlow Datasets. It standardizes data as episodes of timesteps carrying observations, actions, and rewards with arbitrary metadata, and its adoption by the Open X-Embodiment collaboration made it a de facto interchange format for large-scale cross-robot demonstration data. Why it matters for physical AI: Pooling robot data across labs and embodiments requires a common episodic schema, and RLDS supplied the standard that let dozens of datasets merge into the corpora behind generalist robot policies. ### RoboCasa Category: Simulation URL: https://almond.bot/glossary/robocasa RoboCasa is a large-scale simulation framework for training generalist home robots, released in 2024, centered on 120 procedurally varied kitchen scenes populated with thousands of high-quality 3D objects, many created with generative AI tools. Built on RoboSuite, it defines 100 everyday tasks with language instructions and scales human demonstrations into large training sets via automated trajectory generation, with support for mobile manipulators and humanoid embodiments. Why it matters for physical AI: Household-scale diversity in simulation is a proving ground for the scaling hypothesis in robotics, and RoboCasa is one of the principal open testbeds for studying whether synthetic data breadth translates into real-world generality. ### RoboCat Category: Foundation Models URL: https://almond.bot/glossary/robocat RoboCat is a self-improving generalist manipulation agent from Google DeepMind, introduced in 2023 and built on the Gato-style decision transformer architecture with a VQ-GAN image tokenizer. Trained across multiple real and simulated arm embodiments, it adapts to new tasks or robots from as few as 100 to 1,000 demonstrations, then generates its own rollout data with the fine-tuned policy to grow the training set for subsequent, more capable generations. Why it matters for physical AI: RoboCat provided early evidence for the self-improvement flywheel, in which a generalist policy's own deployments produce the data that trains its successor, a loop central to scaling robot foundation models. ### RoboMimic Category: Data & Benchmarks URL: https://almond.bot/glossary/robomimic RoboMimic is an open-source framework and benchmark for robot learning from demonstrations, introduced by Mandlekar et al. in 2021 alongside a systematic study of imitation learning on human demonstration data of varying quality. It provides standardized datasets, including proficient-human and mixed-quality multi-human collections across simulated manipulation tasks, together with reference implementations of algorithms such as behavior cloning, recurrent BC, and offline RL baselines. Why it matters for physical AI: RoboMimic quantified how demonstration quality and quantity drive policy performance, and its datasets and training recipes became reference infrastructure for developing and comparing imitation learning methods. ### RoboSuite Category: Simulation URL: https://almond.bot/glossary/robosuite RoboSuite is a modular, MuJoCo-based simulation framework for robot manipulation research that composes benchmark environments from standardized robots, including Panda, Sawyer, and bimanual configurations, along with interchangeable grippers, controllers, and task arenas. It emphasizes reproducible benchmarking with realistic controller stacks such as operational space control, and it serves as the simulation substrate for RoboMimic and RoboCasa. Why it matters for physical AI: Standardized simulated platforms with realistic controllers keep manipulation research comparable across labs, and RoboSuite's modularity made it a common foundation for demonstration datasets and benchmark suites. ### Robot Cable Category: Hardware URL: https://almond.bot/glossary/robot-cable A robot cable is cabling engineered to survive the continuous flexing, twisting, and torsion of robot motion, carrying power, servo feedback, network signals, and end-effector services along the arm in a managed package called a dresspack. High-flex designs use fine-stranded conductors, torsion-rated shielding, and abrasion-resistant jackets validated for millions of cycles, and many modern arms route cabling internally to eliminate snagging and simplify the work envelope. Why it matters for physical AI: Cable fatigue is one of the most common wear-out failures in deployed robot fleets, and dressing constraints materially affect reachable workspace, tooling design, and maintenance schedules. ### Robot Controller Category: Hardware URL: https://almond.bot/glossary/robot-controller A robot controller is the computer and power-electronics unit that runs a robot's motion control, executing trajectory interpolation, closing joint servo loops at kilohertz rates, and managing safety functions, I/O, and program execution. Industrial controllers pair a cabinet of servo drives with a teach pendant interface, and most modern units expose external real-time interfaces that let outboard computers, including those running learned policies, stream setpoints. Why it matters for physical AI: The controller's external interfaces, streaming rates, and safety architecture define how learned policies can command an arm, making controller openness a practical gatekeeper for deploying physical AI on industrial hardware. ### Robot Nurse Category: Applications URL: https://almond.bot/glossary/robot-nurse A robot nurse is a robotic system that assists nursing staff in healthcare settings, a category spanning hospital logistics robots such as Moxi and the TUG that fetch supplies and deliver medications, patient-transfer research platforms like RIBA designed to lift patients, and telepresence units for remote clinical checks. Deployed systems focus on non-clinical support tasks, relieving documented nursing shortages rather than replacing direct patient care. Why it matters for physical AI: Hospitals combine cluttered human spaces, strict safety and hygiene requirements, and around-the-clock duty cycles, making nursing support a demanding early market for mobile manipulation among people. ### Robot Utility Models (RUM) Category: Foundation Models URL: https://almond.bot/glossary/robot-utility-models Robot Utility Models (RUMs) are pretrained task-specific manipulation policies, introduced by NYU and Hello Robot researchers in 2024, designed to deploy zero-shot in previously unseen environments without any fine-tuning. Each model is trained on roughly a thousand demonstrations gathered across dozens of real environments with the Stick, an inexpensive smartphone-based handheld data collection tool, and deployment adds a vision-language-model success verifier that triggers retries, reaching roughly 90 percent success on tasks such as opening doors and drawers. Why it matters for physical AI: RUMs put concrete numbers on the data diversity needed for zero-shot generalization and showed that cheap handheld collection plus self-verification can yield deployable skills without on-site training. ### Robust Control Category: Control URL: https://almond.bot/glossary/robust-control Robust control is the branch of control theory that designs controllers guaranteeing stability and performance for every plant within a bounded uncertainty set, rather than for one nominal model. Core tools include H-infinity synthesis, mu-analysis for structured uncertainty, and sliding mode control's invariance to matched disturbances. It contrasts with adaptive control, which estimates uncertain parameters online instead of guarding against their worst case. Why it matters for physical AI: Payload changes, joint friction drift, and contact variability mean robot models are always wrong to some degree, and robust design principles inform everything from joint servos to certifying learned controllers under disturbance bounds. ### Robust Estimation Category: Perception URL: https://almond.bot/glossary/robust-estimation Robust estimation is the family of estimation techniques that tolerate outliers and non-Gaussian noise, including M-estimators with influence-limiting loss functions such as Huber and Cauchy, RANSAC-style consensus methods, and outlier-rejecting formulations in factor-graph optimization like switchable constraints and graduated non-convexity. In robot perception it protects pose estimation, registration, and SLAM back ends from mismatched features and spurious loop closures. Why it matters for physical AI: A single unrejected outlier can corrupt an entire map or pose estimate, so robust losses and consensus checks are what keep state estimation trustworthy on real, messy sensor data. ### Robustness to Disturbances Category: Control URL: https://almond.bot/glossary/robustness-to-disturbances Robustness to disturbances is a robot system's ability to maintain stability and task performance under external perturbations such as pushes, collisions, payload changes, terrain irregularities, and sensor dropouts. It is achieved through feedback control margins, robust and adaptive design, and, in learned systems, by training across randomized perturbations so recovery behavior is baked into the policy, and it is measured with standardized push, slip, and load tests. Why it matters for physical AI: Deployment environments never match training or design conditions exactly, and disturbance robustness is frequently the deciding gap between an impressive demo and a robot that survives contact with the real world. ### ROS (Robot Operating System) Category: Software & Middleware URL: https://almond.bot/glossary/ros ROS (Robot Operating System) is an open-source robotics middleware, begun at Willow Garage in 2007, that structures robot software as a graph of nodes communicating through publish-subscribe topics, services, and actions, with standardized message types and a vast package ecosystem covering drivers, perception, and planning. Despite its name it is not an operating system, and the original ROS 1 line ended official support with Noetic's end of life in May 2025. Why it matters for physical AI: ROS created the shared conventions and reusable components that let robotics teams avoid rebuilding infrastructure, and its interfaces remain the lingua franca for integrating sensors, arms, and learned modules. ### ROS 2 Category: Software & Middleware URL: https://almond.bot/glossary/ros-2 ROS 2 is the ground-up redesign of the Robot Operating System, replacing ROS 1's custom transport and central master with the DDS publish-subscribe standard for discovery and communication. It adds configurable quality-of-service policies, support for real-time and embedded systems via micro-ROS, Windows and macOS support, security through SROS2, and managed node lifecycles, with long-term-support distributions such as Humble and Jazzy targeting production and commercial deployment. Why it matters for physical AI: Production robot fleets need deterministic communication, security, and multi-robot support that research-era middleware lacked, and ROS 2 is the industry's consensus path for building such systems on open source. ### ROS Bag Category: Software & Middleware URL: https://almond.bot/glossary/ros-bag A ROS bag is a recording of timestamped ROS messages captured from live topics, replayable later to reproduce a robot's exact sensor and state streams for debugging, algorithm development, and dataset creation. The rosbag2 implementation in ROS 2 supports pluggable storage backends, with the MCAP container format now the default. Bags are the raw material from which many robot learning datasets are curated and converted. Why it matters for physical AI: Log-and-replay is the foundation of robotics debugging and data engineering, and bag archives from deployed fleets are increasingly mined as training corpora for learned perception and control. ### Rotation Matrix Category: Math & Kinematics URL: https://almond.bot/glossary/rotation-matrix A rotation matrix is a 3x3 orthogonal matrix with determinant +1 that represents orientation or rotates vectors in three-dimensional space, with the set of all such matrices forming the special orthogonal group SO(3). It uses nine numbers for three degrees of freedom but composes by multiplication, avoids the gimbal-lock singularities of Euler angles, and requires re-orthonormalization under numerical drift; quaternions offer a more compact alternative. Why it matters for physical AI: Orientation representation is a daily engineering decision in robotics, affecting interpolation, control error definitions, and network output parameterizations, where learned models often favor continuous rotation representations over Euler angles. ### RT-1 Category: Foundation Models URL: https://almond.bot/glossary/rt-1 RT-1 (Robotics Transformer 1) is a 35-million-parameter transformer policy from Google, introduced in 2022, trained on roughly 130,000 real-world demonstrations spanning more than 700 tasks, collected over 17 months with a fleet of 13 mobile manipulators. It encodes images with a FiLM-conditioned EfficientNet, compresses tokens with TokenLearner, and outputs discretized arm and base actions at 3 Hz, demonstrating that large, diverse robot datasets yield broad task generalization. Why it matters for physical AI: RT-1 was the proof of concept that scaling diverse real-robot data with a high-capacity architecture improves generalization, setting the data-scaling agenda that led to RT-2 and cross-embodiment training. ### RT-2 Category: Foundation Models URL: https://almond.bot/glossary/rt-2 RT-2 is a vision-language-action model from Google DeepMind, introduced in 2023, created by co-fine-tuning pretrained vision-language models, PaLI-X and PaLM-E, on robot trajectories with actions expressed as text tokens in the model's existing vocabulary. Co-training on web data alongside robot data transferred semantic knowledge to control, yielding emergent capabilities such as reasoning about novel object categories, icons, and simple symbolic instructions, with variants up to 55 billion parameters. Why it matters for physical AI: RT-2 defined the vision-language-action recipe now standard across the field: start from a web-pretrained multimodal backbone and treat robot actions as another language the model learns to speak. ### RT-X Category: Foundation Models URL: https://almond.bot/glossary/rt-x RT-X refers to the cross-embodiment robot policies, RT-1-X and RT-2-X, trained on the Open X-Embodiment dataset, a 2023 collaboration of more than 20 institutions that pooled over one million real robot trajectories from 22 embodiments into a unified format. The RT-X models outperformed the original models trained on each lab's own data, providing the first large-scale evidence of positive transfer across different robot platforms. Why it matters for physical AI: Positive cross-embodiment transfer is the empirical bedrock of the generalist-robot-policy agenda, indicating that data pooled across heterogeneous robots compounds rather than conflicts. ### RViz Category: Software & Middleware URL: https://almond.bot/glossary/rviz RViz is the standard 3D visualization tool for ROS, rendering live sensor streams, coordinate transform trees, robot models from URDF descriptions, point clouds, occupancy maps, camera overlays, and planning scenes in an interactive viewport. Interactive markers allow direct manipulation of poses and goals, making it both a debugging instrument and a lightweight operator interface, with RViz2 providing the equivalent for ROS 2. Why it matters for physical AI: Seeing what the robot believes, including its frames, sensor returns, and planned motions, is how developers diagnose perception and planning failures, and RViz is the field's shared window into that internal state. ## Glossary: S ### Safety Constraints in Robotics Category: Safety & Standards URL: https://almond.bot/glossary/safety-constraints-in-robotics Safety constraints in robotics are explicit limits on robot behavior (joint position and velocity bounds, force and torque caps, workspace boundaries, and human-proximity rules) that a system must satisfy at all times during operation. In learning-based control, they are often formalized as constrained Markov decision processes (CMDPs), where a policy maximizes reward subject to bounds on expected constraint violations. Enforcement mechanisms range from hard-coded software limits to control barrier functions and runtime shielding. Why it matters for physical AI: Learned policies offer no inherent guarantees, so explicit constraint enforcement is what makes it acceptable to run neural network controllers on physical hardware around people and property. ### Safety Layer Category: Safety & Standards URL: https://almond.bot/glossary/safety-layer A safety layer is a runtime module that sits between a controller or learned policy and the robot's actuators, filtering or overriding commands that would violate safety constraints. Implementations include control barrier function-based quadratic program filters, action shielding, velocity scaling near humans, and hard emergency-stop logic. The layer is typically designed and verified independently of the policy it supervises, so guarantees hold regardless of policy behavior. Why it matters for physical AI: Wrapping an unverifiable neural policy in a verifiable safety filter is the dominant pattern for deploying robot foundation models on real hardware without certifying the model itself. ### Safety-Rated Monitored Stop Category: Safety & Standards URL: https://almond.bot/glossary/safety-rated-monitored-stop Safety-rated monitored stop is one of the four collaborative operation modes defined in ISO 10218 and ISO/TS 15066, in which the robot halts motion whenever a human enters the collaborative workspace but keeps its drives energized. Safety-rated sensors monitor the stop condition, and motion may resume automatically once the person leaves. It allows a human and robot to share a workspace sequentially without a full protective shutdown and restart cycle. Why it matters for physical AI: Understanding the standardized collaborative modes is essential when deploying manipulation systems in shared human workspaces, since certification and risk assessment hinge on which mode the cell implements. ### Sample Efficiency Category: Robot Learning URL: https://almond.bot/glossary/sample-efficiency Sample efficiency is the amount of data or environment interaction a learning algorithm needs to reach a given level of performance. In robotics it is a first-order concern because real-world data collection is slow, hardware wears out, and failures can be destructive. Off-policy reinforcement learning, demonstration bootstrapping, pretrained representations, and simulation are the standard levers for improving it. Why it matters for physical AI: Real robot time is the scarcest resource in physical AI; methods that learn from hundreds rather than millions of trials determine what is practical to train outside of simulation. ### Sampling-Based Motion Planning Category: Math & Kinematics URL: https://almond.bot/glossary/sampling-based-motion-planning Sampling-based motion planning is a family of algorithms that find collision-free paths by randomly sampling a robot's configuration space and connecting valid samples into a graph or tree, rather than discretizing the space exhaustively. Probabilistic roadmaps (PRM, Kavraki et al., 1996) and rapidly-exploring random trees (RRT, LaValle, 1998) are the canonical methods, with asymptotically optimal variants like RRT* following later. These planners scale to the high-dimensional configuration spaces of manipulator arms where grid-based search is intractable. Why it matters for physical AI: Even as end-to-end policies advance, sampling-based planners remain the workhorse for collision-free arm motion, and hybrid systems routinely delegate free-space transit to them while policies handle contact. ### SayCan Category: Foundation Models URL: https://almond.bot/glossary/saycan SayCan is a method from Google Research ("Do As I Can, Not As I Say", 2022) that grounds large language model planning in robot capabilities by combining the LLM's estimate of a skill's usefulness with a learned value function estimating the skill's probability of success in the current state. The system executed long-horizon kitchen tasks on mobile manipulators by iteratively selecting the skill with the highest combined score. It was an early and influential demonstration of LLMs as robot task planners. Why it matters for physical AI: SayCan established the now-standard pattern of pairing internet-scale language reasoning with grounded affordance estimates, a lineage that runs directly to today's vision-language-action models. ### Scaling Laws for Robot Learning Category: Robot Learning URL: https://almond.bot/glossary/scaling-laws-for-robot-learning Scaling laws for robot learning are empirical relationships describing how policy performance improves with the amount of training data, model capacity, and data diversity. Unlike language modeling, where loss follows smooth power laws in tokens and parameters, robotics results emphasize diversity: studies on datasets like Open X-Embodiment suggest that generalization scales with the variety of environments, objects, and embodiments rather than raw demonstration count alone. Establishing reliable scaling laws remains an open research question. Why it matters for physical AI: Whether robot policies scale predictably with data determines the economics of the entire field, guiding decisions about teleoperation fleets, simulation, and cross-embodiment data pooling. ### SCARA Robot Category: Hardware URL: https://almond.bot/glossary/scara-robot A SCARA robot (Selective Compliance Assembly Robot Arm) is a four-axis manipulator with two parallel revolute joints providing motion in the horizontal plane, a vertical linear axis, and a wrist rotation, making it rigid vertically but compliant laterally. Invented by Hiroshi Makino at Yamanashi University in 1981, the architecture excels at fast, repeatable pick-and-place and vertical insertion tasks. SCARAs remain among the fastest and most precise arms for planar assembly work. Why it matters for physical AI: High-speed electronics assembly and kitting lines still favor SCARAs over six-axis arms, so learned policies targeting these domains must handle their restricted, planar workspace geometry. ### Scene Graph Category: Perception URL: https://almond.bot/glossary/scene-graph A scene graph is a structured representation of a visual scene as a graph whose nodes are objects (with attributes) and whose edges encode pairwise relationships such as "on top of", "inside", or "attached to". Scene graph generation from images is an established computer vision task, and 3D scene graphs extend the idea to hierarchical maps spanning objects, places, rooms, and buildings. Robots use scene graphs as symbolic scaffolding for task planning and spatial queries. Why it matters for physical AI: Scene graphs bridge raw perception and symbolic reasoning, giving language model planners a queryable world state and enabling long-horizon tasks that pure pixel-to-action policies handle poorly. ### Screw Theory Category: Math & Kinematics URL: https://almond.bot/glossary/screw-theory Screw theory is a mathematical framework describing rigid body motion and forces using screws: any spatial displacement can be expressed as a rotation about an axis combined with a translation along it (Chasles' theorem), and any system of forces reduces to a force plus a couple along an axis (Poinsot). Velocities become twists and forces become wrenches, both six-dimensional vectors. It underpins the product-of-exponentials formulation of kinematics popularized by texts like Lynch and Park's Modern Robotics. Why it matters for physical AI: Twists and wrenches provide the coordinate-free language in which modern manipulator kinematics, dynamics, and contact models are written, and they surface directly in software like screw-based IK solvers. ### SE(3) Category: Math & Kinematics URL: https://almond.bot/glossary/se3 SE(3) is the special Euclidean group in three dimensions, the set of all rigid body transformations combining a rotation in SO(3) with a translation in R^3. Every robot pose (of a link, end-effector, camera, or object) is an element of SE(3), commonly represented as a 4x4 homogeneous transformation matrix, and its Lie algebra se(3) represents twists. The group structure (composition, inversion, exponential and logarithm maps) is fundamental to kinematics, state estimation, and pose interpolation. Why it matters for physical AI: Robot learning increasingly exploits SE(3) structure directly, from equivariant grasp networks to policies that output end-effector poses, making fluency with the group a baseline requirement. ### Search and Rescue Robot Category: Applications URL: https://almond.bot/glossary/search-and-rescue-robot A search and rescue robot is a platform designed to locate victims and assess hazards in disaster environments (collapsed structures, mines, floods, or nuclear incidents) that are too dangerous or inaccessible for human responders. The field spans tracked crawlers, snake robots, aerial drones, and legged systems, and has been shaped by competitions like the DARPA Robotics Challenge and DARPA Subterranean Challenge. Robots were deployed at events including the World Trade Center collapse and the Fukushima Daiichi accident. Why it matters for physical AI: Disaster response stresses every hard problem at once (degraded sensing, unreliable communication, unstructured terrain, and human trust), making it a proving ground for autonomy that generalizes. ### Segment Anything (SAM) Category: Perception URL: https://almond.bot/glossary/segment-anything Segment Anything (SAM) is a promptable image segmentation foundation model released by Meta AI in 2023, trained on the SA-1B dataset of over one billion masks, that produces object masks from point, box, or mask prompts and generalizes zero-shot to unseen domains. SAM 2 (2024) extended the model to video with streaming memory for real-time mask tracking. Robotics pipelines widely use SAM to segment target objects for grasping, tracking, and data annotation without task-specific training. Why it matters for physical AI: Off-the-shelf, promptable segmentation removed a whole category of perception engineering from robot pipelines, letting manipulation systems isolate arbitrary objects with a click or a language-grounded detector. ### Self-Play Category: Robot Learning URL: https://almond.bot/glossary/self-play Self-play is a reinforcement learning technique in which an agent improves by training against copies or past versions of itself, generating an automatic curriculum whose difficulty scales with the agent's own competence. It powered landmark results such as AlphaGo, AlphaZero, and OpenAI Five. In robotics, related asymmetric variants pit a goal-proposing agent against a goal-solving agent, or an adversary that perturbs dynamics against a robust policy. Why it matters for physical AI: Automatic curricula from self-play sidestep manual task design, and adversarial variants are a practical route to policies robust enough for hardware deployment. ### Self-Supervised Learning Category: Robot Learning URL: https://almond.bot/glossary/self-supervised-learning Self-supervised learning is a paradigm in which models learn representations from unlabeled data by solving pretext tasks whose labels are derived from the data itself, such as contrastive instance discrimination, masked reconstruction, or temporal prediction. Vision encoders like DINOv2 and robotics-oriented representations like R3M (trained on the Ego4D human video corpus) are self-supervised. Robots can also self-supervise directly, for example labeling grasp outcomes from their own trial data. Why it matters for physical AI: Unlabeled video and autonomous robot experience vastly outnumber curated demonstrations, so self-supervision is the main mechanism for turning cheap data into pretrained perception and dynamics representations. ### Semantic 3D Reconstruction Category: Perception URL: https://almond.bot/glossary/semantic-3d-reconstruction Semantic 3D reconstruction is the joint recovery of a scene's 3D geometry and per-element semantic labels, producing maps whose surfaces, voxels, or points carry object categories or open-vocabulary embeddings. Classical pipelines fuse depth into TSDF or voxel grids while projecting 2D segmentation into the volume; recent systems distill CLIP-style features into 3D fields so maps can be queried with natural language. The output supports object-aware planning rather than purely geometric collision avoidance. Why it matters for physical AI: Language-queryable 3D maps let robots ground instructions like "fetch the mug by the sink" in metric space, connecting foundation model semantics to executable motion. ### Semantic Grasping Category: Manipulation URL: https://almond.bot/glossary/semantic-grasping Semantic grasping is grasp selection conditioned on what an object is and how it will be used, rather than on geometry and stability alone: grasping a knife by the handle for handover, or a mug away from its rim for pouring. Approaches combine object recognition, part segmentation, and affordance prediction with grasp synthesis, and recent systems condition grasp generation on natural language task descriptions. It contrasts with purely analytic or data-driven stable-grasp pipelines. Why it matters for physical AI: Household and industrial tasks rarely want just any stable grasp; task-appropriate, instruction-following grasp choice is a prerequisite for general-purpose manipulation in human environments. ### Semantic Segmentation Category: Perception URL: https://almond.bot/glossary/semantic-segmentation Semantic segmentation is the pixel-wise classification of an image into semantic categories, assigning every pixel a label such as table, robot, or floor without distinguishing object instances. Fully convolutional networks (Long et al., 2015) established the modern approach, followed by architectures like DeepLab, U-Net, and transformer-based SegFormer and Mask2Former. Instance and panoptic segmentation extend the task to separate individual objects. Why it matters for physical AI: Dense per-pixel semantics feed nearly every downstream robotic capability (traversability estimation, object-aware mapping, grasp target isolation) and remain a standard intermediate representation in modular perception stacks. ### Semantic SLAM Category: Navigation & SLAM URL: https://almond.bot/glossary/semantic-slam Semantic SLAM is simultaneous localization and mapping augmented with semantic information, producing maps that contain labeled objects and regions rather than only geometry. Systems detect and segment objects during mapping, insert them as landmarks in the factor graph, and can exploit semantics for data association and loop closure: recognizing a place by its objects rather than raw appearance. Object-level SLAM systems and 3D scene graph builders are representative examples. Why it matters for physical AI: Maps that know what things are, not just where surfaces lie, are the substrate for instruction following, mobile manipulation, and long-term autonomy in changing environments. ### Sensor Fusion Category: Perception URL: https://almond.bot/glossary/sensor-fusion Sensor fusion is the combination of measurements from multiple sensors into a single, more accurate and reliable estimate than any individual source provides. Classical robotics fuses IMU, wheel odometry, LiDAR, GPS, and cameras through probabilistic estimators such as Kalman filters and factor graph optimization; visual-inertial odometry is a canonical example. In learned systems, fusion also happens inside networks that ingest multimodal streams like RGB, depth, proprioception, and touch. Why it matters for physical AI: No single sensor survives all conditions (cameras fail in darkness, IMUs drift, LiDAR struggles with glass), so principled fusion underwrites the state estimates every controller depends on. ### Sensor-Based Grasping Category: Manipulation URL: https://almond.bot/glossary/sensor-based-grasping Sensor-based grasping is grasp execution guided by real-time sensor feedback (vision, tactile arrays, force-torque sensing, or proximity sensors) rather than open-loop replay of a preplanned grasp pose. Closed-loop approaches adjust the approach and finger closure as new observations arrive, correcting for perception error, object movement, and slip. QT-Opt (2018), which learned closed-loop vision-based grasping from over 500,000 real grasp attempts, is a landmark example. Why it matters for physical AI: Open-loop grasps fail whenever calibration drifts or objects shift; feedback-driven grasping is what makes picking robust in cluttered, dynamic settings like bins and homes. ### Sensorimotor Learning Category: Robot Learning URL: https://almond.bot/glossary/sensorimotor-learning Sensorimotor learning is the acquisition of coupled perception-action skills, where a robot learns to map raw sensory streams directly to motor commands rather than passing through hand-designed intermediate representations. End-to-end visuomotor policy training (Levine et al., 2016) demonstrated that convolutional networks could learn such mappings on physical manipulators. The term borrows from neuroscience, where it describes how animals refine movement through sensory feedback. Why it matters for physical AI: Tight perception-action coupling captures contact-rich subtleties that modular pipelines discard, and it is the operating principle behind modern visuomotor policies and vision-language-action models. ### Serial Robot Category: Hardware URL: https://almond.bot/glossary/serial-robot A serial robot is a manipulator built as an open kinematic chain, with links connected end to end by joints from a fixed base to the end-effector, so each actuator carries all the links and motors distal to it. Most industrial six-axis arms and SCARA robots are serial. The architecture offers a large workspace and simple kinematics relative to parallel mechanisms, at the cost of lower stiffness and accumulated joint errors along the chain. Why it matters for physical AI: Serial arms dominate the platforms on which manipulation policies are trained and deployed, and their cantilevered structure shapes payload, compliance, and calibration behavior that learned controllers must absorb. ### Series Elastic Actuator (SEA) Category: Hardware URL: https://almond.bot/glossary/series-elastic-actuator A series elastic actuator (SEA) is an actuator that deliberately places a compliant element, typically a spring, between the motor's gear train and the load, introduced by Pratt and Williamson at MIT in 1995. Measuring spring deflection turns force control into a position measurement problem, yielding accurate force output, shock tolerance, and low mechanical output impedance. SEAs have been used in legged robots, exoskeletons, and collaborative arms such as Rethink's Baxter. Why it matters for physical AI: Compliant actuation makes sustained, safe physical contact tractable, which is why SEA-style designs recur wherever robots must touch people or absorb impacts rather than track trajectories rigidly. ### SERL Category: Robot Learning URL: https://almond.bot/glossary/serl SERL (Sample-Efficient Robotic reinforcement Learning) is an open-source software suite from UC Berkeley (Luo et al., 2024) for training reinforcement learning policies directly on real manipulators, combining a high update-to-data ratio off-policy learner seeded with demonstrations, learned reward classifiers, and automatic environment resets. It demonstrated tasks like PCB insertion and cable routing learned in tens of minutes of real interaction. The HIL-SERL follow-up added human-in-the-loop corrections, reaching near-perfect success on contact-rich tasks. Why it matters for physical AI: SERL made real-world RL reproducible for ordinary labs, showing that on-robot reinforcement learning can polish policies past the ceiling of pure imitation on precision tasks. ### Servo Motor Category: Hardware URL: https://almond.bot/glossary/servo-motor A servo motor is a motor packaged with feedback sensing and closed-loop control so it can track commanded position, velocity, or torque. The category spans inexpensive hobby servos with integrated potentiometers and controllers, smart serial-bus servos like Dynamixel and Feetech units common in research arms, and industrial AC servos with high-resolution encoders and external drives. Feedback control distinguishes servos from open-loop actuators such as steppers. Why it matters for physical AI: Nearly every joint on a modern manipulator is a servo of some kind, and the quality of its torque control and feedback resolution bounds what any learned controller can achieve. ### Shadow Hand Category: Hardware URL: https://almond.bot/glossary/shadow-hand The Shadow Hand is an anthropomorphic dexterous robotic hand from the Shadow Robot Company, with 24 joints and roughly 20 actuated degrees of freedom driven by tendons, closely matching the kinematics of a human hand. It became the reference platform for dexterous manipulation research, most prominently in OpenAI's Learning Dexterity work (2018-2019), where policies trained in simulation with domain randomization performed in-hand cube reorientation and Rubik's cube manipulation on the physical hand. Why it matters for physical AI: Human-equivalent hands expose the frontier of contact-rich control; results on platforms like the Shadow Hand calibrate how far sim-to-real training can carry high-dimensional dexterity. ### Shared Autonomy Category: Teleoperation URL: https://almond.bot/glossary/shared-autonomy Shared autonomy is a control paradigm in which a human operator and an autonomous system jointly control a robot, with the autonomy inferring operator intent and blending or arbitrating between human input and assistive commands. Formulations range from linear command blending to POMDP-based intent inference over candidate goals. It is applied in assistive manipulation, surgical robotics, and remote operation over degraded communication links. Why it matters for physical AI: Full autonomy fails on edge cases and full teleoperation does not scale, so blending the two is the pragmatic deployment path, and the corrections it generates become training data. ### Sim-to-Real Gap Category: Simulation URL: https://almond.bot/glossary/sim-to-real-gap The sim-to-real gap is the discrepancy between a robot's behavior in simulation and on physical hardware, arising from unmodeled dynamics, simplified contact and friction models, actuator nonlinearities, sensing artifacts, latency, and visual differences between rendered and real observations. Policies that exploit simulator quirks often fail outright when transferred. The gap is attacked from both sides: better system identification and rendering narrow it, while domain randomization trains policies to be insensitive to it. Why it matters for physical AI: Every simulation-trained policy pays this tax at deployment; quantifying and shrinking the gap decides whether cheap simulated experience translates into real-world competence. ### Sim-to-Real Transfer Category: Simulation URL: https://almond.bot/glossary/sim-to-real-transfer Sim-to-real transfer is the practice of training robot policies in simulation and deploying them on physical hardware, exploiting simulation's speed, safety, and parallelism while confronting the mismatch between simulated and real dynamics and sensing. Standard techniques include domain randomization (Tobin et al., 2017), system identification, teacher-student distillation from privileged simulator state, and real-world fine-tuning. Landmark successes include OpenAI's dexterous in-hand manipulation and learned locomotion controllers for ANYmal and other legged robots. Why it matters for physical AI: Simulation offers effectively unlimited experience at near-zero marginal cost; transfer techniques determine how much of that experience survives contact with reality, especially for locomotion and dexterity. ### Simulation Benchmark Category: Data & Benchmarks URL: https://almond.bot/glossary/simulation-benchmark A simulation benchmark is a standardized suite of simulated environments, tasks, and evaluation protocols used to compare robot learning methods reproducibly. Prominent examples include Meta-World and RLBench for manipulation, LIBERO for lifelong and language-conditioned learning, ManiSkill and Isaac Lab for GPU-parallel evaluation, and Habitat for embodied navigation. Benchmarks fix observation spaces, success criteria, and train-test splits so reported numbers are comparable across papers. Why it matters for physical AI: Reproducible comparison is chronically hard on physical robots, so simulated benchmarks carry most of the field's empirical burden, while their gap to real-world difficulty must be kept in mind. ### Singularity Category: Math & Kinematics URL: https://almond.bot/glossary/singularity A singularity is a robot configuration at which the manipulator Jacobian loses rank, so the end-effector instantaneously loses the ability to move or exert force in one or more directions. Near singularities, small Cartesian velocities demand unbounded joint velocities, causing erratic high-speed joint motion. Six-axis arms exhibit characteristic wrist, shoulder, and elbow singularities, such as the wrist singularity when joints four and six align. Why it matters for physical AI: Policies and planners that command Cartesian motion must respect singular regions, since crossing one can trigger joint-speed faults or violent arm motion on real hardware. ### Singularity Avoidance Category: Control URL: https://almond.bot/glossary/singularity-avoidance Singularity avoidance is the set of techniques for keeping a manipulator away from, or behaving stably near, singular configurations. Common methods include damped least-squares (Levenberg-Marquardt) inverse kinematics that trades tracking accuracy for bounded joint velocities, manipulability-maximizing null-space motion for redundant arms, and planners that treat low-manipulability regions as obstacles. Singularity-robust inverse kinematics is standard in industrial and research control stacks. Why it matters for physical AI: Learned policies that output end-effector targets rely on the underlying IK layer to survive singular regions, making robust inversion a quiet prerequisite for safe deployment. ### Skill Chaining Category: Robot Learning URL: https://almond.bot/glossary/skill-chaining Skill chaining is the composition of learned skills into longer behaviors by ensuring each skill terminates in a state from which the next skill can succeed, formally, within the next skill's initiation set. The idea originates in the options framework of hierarchical reinforcement learning (Konidaris and Barto, 2009) and reappears in long-horizon manipulation, where policies are trained or regularized so their terminal state distributions match successors' initial state distributions. Why it matters for physical AI: Long-horizon tasks like cooking or assembly exceed what monolithic policies learn reliably; chaining verified skills with compatible boundary conditions is a scalable route to multi-step competence. ### Skill Library Category: Robot Learning URL: https://almond.bot/glossary/skill-library A skill library is a curated collection of reusable robot skills (learned policies, motion primitives, or parameterized controllers), each with defined inputs, preconditions, and effects, from which a planner or language model selects and sequences behaviors at run time. Systems like SayCan plan directly over such libraries, and lifelong learning research studies growing libraries autonomously by acquiring and naming new skills. Libraries trade end-to-end flexibility for modularity, verification, and reuse. Why it matters for physical AI: Foundation model planners are only as capable as the skills they can invoke; a well-tested library turns open-ended language instructions into executable, auditable robot behavior. ### Skill Primitive Category: Robot Learning URL: https://almond.bot/glossary/skill-primitive A skill primitive is a basic, parameterized unit of robot behavior (such as grasp, push, pour, or move-to-pose) that accepts arguments like target poses or objects and can be composed into complex tasks. Primitives may be implemented as hand-coded controllers, dynamic movement primitives, or small learned policies. Hierarchical systems act over primitives instead of raw joint commands, shortening effective task horizons by orders of magnitude. Why it matters for physical AI: Choosing the action abstraction is a central design decision in robot learning: primitives dramatically improve sample efficiency and planning tractability at some cost in dexterity and expressiveness. ### SLAM Category: Navigation & SLAM URL: https://almond.bot/glossary/slam SLAM (Simultaneous Localization and Mapping) is the problem of building a map of an unknown environment while concurrently estimating the robot's pose within it, using sensors such as LiDAR, cameras, and IMUs. Solutions evolved from EKF-SLAM and particle filters to modern graph-based optimization over poses and landmarks, with loop closure correcting accumulated drift. ORB-SLAM, LIO-SAM, and Cartographer are widely used systems. Why it matters for physical AI: Mobile robots cannot presume known maps or external positioning; SLAM supplies the spatial self-awareness on which navigation, exploration, and mobile manipulation are built. ### Sliding Mode Control Category: Control URL: https://almond.bot/glossary/sliding-mode-control Sliding mode control is a nonlinear, variable-structure control technique that drives the system state onto a designed sliding surface in finite time and keeps it there with a discontinuous switching control law. Once on the surface, the closed-loop dynamics are governed by the surface definition and are insensitive to matched disturbances and model uncertainty. Its main drawback is chattering (high-frequency switching that excites unmodeled dynamics) mitigated by boundary layers and higher-order sliding modes. Why it matters for physical AI: Strong robustness to model error with modest computation keeps sliding mode control relevant for underwater vehicles, quadrotors, and manipulators whose dynamics are poorly identified. ### Sliding Surface Category: Control URL: https://almond.bot/glossary/sliding-surface A sliding surface is the manifold in state space, defined by a switching function s(x) = 0, on which a sliding mode controller constrains the system to evolve. The surface is designed so that motion restricted to it exhibits the desired stable dynamics: for example, s = de + Le forces tracking error e to decay exponentially. Controller design then reduces to guaranteeing finite-time reaching and invariance of the surface. Why it matters for physical AI: Surface design is where control objectives are encoded in sliding mode methods, translating tracking specifications for arms and vehicles into geometric constraints with provable convergence. ### Slip Ring Category: Hardware URL: https://almond.bot/glossary/slip-ring A slip ring is an electromechanical component that transmits electrical power and signals across a rotating interface using conductive rings and sliding brushes or contacts, allowing a joint to rotate continuously without twisting cables. Slip rings appear in spinning LiDAR units, pan-tilt turrets, robot wrists, and cable reels; contactless variants pass data inductively or optically to avoid brush wear. Bandwidth, channel count, and contact noise are the key specifications. Why it matters for physical AI: Cable management quietly limits joint range on many arms; continuous-rotation joints enabled by slip rings remove wind-up constraints that otherwise complicate motion planning and teleoperation. ### Smart Factory Category: Applications URL: https://almond.bot/glossary/smart-factory A smart factory is a manufacturing facility in which machines, robots, sensors, and software are networked and data-driven, enabling self-monitoring, adaptive scheduling, and highly automated production: the factory-level realization of Industry 4.0. Characteristic elements include autonomous mobile robots for intralogistics, collaborative robot cells, digital twins of lines and processes, and predictive maintenance from streaming sensor data. Why it matters for physical AI: Factories are the leading deployment ground for general-purpose robots, and smart factory infrastructure (connectivity, digital twins, structured data) is precisely what large-scale robot fleet learning requires. ### SmolVLA Category: Foundation Models URL: https://almond.bot/glossary/smolvla SmolVLA is a compact open-source vision-language-action model released by Hugging Face in 2025, with roughly 450 million parameters, designed to run on consumer hardware including single GPUs and CPUs. It pairs a pruned SmolVLM backbone with a flow-matching action expert, was trained on community-contributed LeRobot datasets rather than proprietary fleet data, and supports asynchronous inference that decouples action prediction from execution. Despite its size, it reported performance competitive with much larger VLAs on benchmark manipulation tasks. Why it matters for physical AI: Small, openly trained VLAs democratize robot foundation models, proving that community data and modest compute can yield deployable policies outside industrial labs. ### Snake Robot Category: Hardware URL: https://almond.bot/glossary/snake-robot A snake robot is a hyper-redundant robot composed of many serially linked segments that locomotes through body undulation (lateral undulation, sidewinding, or rolling gaits) rather than wheels or legs. Carnegie Mellon's modular snake robots are a canonical research line, demonstrating pipe climbing, swimming, and pole wrapping. Their slender form suits confined spaces such as pipelines, industrial inspection cavities, and collapsed structures in search and rescue. Why it matters for physical AI: Hyper-redundant bodies reach where conventional platforms cannot, and their many-DOF gait control problems have become fertile testbeds for learned locomotion in constrained contact-rich settings. ### SO-100 Arm Category: Hardware URL: https://almond.bot/glossary/so-100-arm The SO-100 arm is a low-cost, open-source robotic arm developed by The Robot Studio with Hugging Face's LeRobot project, built from 3D-printed parts and Feetech serial-bus servos for a parts cost on the order of one hundred dollars per arm. It is typically used as a leader-follower teleoperation pair for collecting imitation learning data. The refined SO-101 (2025) improved assembly and gearing, and the design became a de facto standard platform for community robot learning. Why it matters for physical AI: Collapsing hardware cost by two orders of magnitude turned data collection and policy training into a hobbyist-accessible activity, seeding the community datasets behind open robot foundation models. ### SO(3) Category: Math & Kinematics URL: https://almond.bot/glossary/so3 SO(3) is the special orthogonal group in three dimensions: the set of all 3x3 rotation matrices, i.e., orthogonal matrices with determinant one, representing every possible orientation of a rigid body. Common parameterizations include unit quaternions, axis-angle vectors, and Euler angles, each with trade-offs around singularities and double cover. Its Lie algebra so(3) consists of skew-symmetric matrices corresponding to angular velocities, connected to the group by the exponential map. Why it matters for physical AI: Orientation representation choices ripple through robot learning: regression targets, interpolation, and equivariant architectures all hinge on handling SO(3)'s topology correctly rather than treating rotations as flat vectors. ### Social Navigation Category: Navigation & SLAM URL: https://almond.bot/glossary/social-navigation Social navigation is robot navigation among people that goes beyond collision avoidance to respect human social conventions: maintaining comfortable distances, not cutting through groups, signaling intent, and yielding appropriately in corridors and doorways. Approaches range from proxemics-informed cost functions and social force models to learned policies trained on pedestrian datasets and evaluated in benchmarks such as SocNavBench. Human motion prediction is a core subproblem. Why it matters for physical AI: Robots entering homes, hospitals, and sidewalks are judged on comfort as much as efficiency; socially compliant motion is a prerequisite for acceptance of mobile robots around people. ### Soft Actor-Critic (SAC) Category: Robot Learning URL: https://almond.bot/glossary/soft-actor-critic Soft Actor-Critic (SAC) is an off-policy, maximum-entropy deep reinforcement learning algorithm (Haarnoja et al., 2018) that trains a stochastic policy to maximize expected return plus policy entropy, using twin Q-networks and a replay buffer. The entropy term encourages exploration and improves robustness, while off-policy reuse of past data yields strong sample efficiency. SAC became a default choice for continuous-control robotics, including learning quadruped locomotion and manipulation skills directly on hardware. Why it matters for physical AI: Sample-efficient, stable off-policy learning is what makes reinforcement learning on physical robots feasible at all, and SAC remains the backbone of most real-world RL pipelines. ### Soft Gripper Category: Hardware URL: https://almond.bot/glossary/soft-gripper A soft gripper is an end-effector that grasps using compliant materials and structures (pneumatic elastomer fingers, fin-ray-effect flexures, or granular jamming bags), which passively conform to object geometry instead of relying on precise finger placement. Compliance provides tolerance to pose uncertainty and gentle handling of fragile or irregular items such as produce and baked goods. Trade-offs include limited payload, imprecise in-hand positioning, and difficulty modeling contact. Why it matters for physical AI: Passive conformity converts perception error into successful grasps for free, which is why soft end-effectors dominate food handling and complement learned grasp policies on hard-to-model objects. ### Soft Robot Category: Hardware URL: https://almond.bot/glossary/soft-robot A soft robot is a robot whose body is made substantially of compliant materials (silicone elastomers, fabrics, or fluid-driven chambers) deforming continuously rather than articulating through discrete rigid joints. Actuation is commonly pneumatic or hydraulic, with cable-driven and shape-memory-alloy variants, and modeling relies on continuum mechanics rather than rigid body kinematics. The field draws heavily on biological inspiration such as octopus arms and elephant trunks. Why it matters for physical AI: Intrinsic compliance offers safety around humans and adaptability to unstructured contact, while the modeling difficulty of continuum bodies makes soft systems a natural target for learned control. ### Space Robot Category: Applications URL: https://almond.bot/glossary/space-robot A space robot is a robotic system operating in orbit or on planetary surfaces, spanning manipulators like the ISS's Canadarm2, planetary rovers such as Curiosity and Perseverance, and free-flying assistants and servicing spacecraft. Distinctive challenges include free-floating dynamics where arm motion disturbs the base, radiation-hardened and thermally extreme hardware, and communication delays from seconds to tens of minutes that force supervisory rather than direct teleoperation. Why it matters for physical AI: Light-minutes of latency make onboard autonomy non-negotiable, so space systems pioneered supervised autonomy, fault tolerance, and long-horizon task execution that terrestrial deployment now borrows. ### Sparse Optical Flow Category: Perception URL: https://almond.bot/glossary/sparse-optical-flow Sparse optical flow is the estimation of apparent image motion at a selected set of feature points across consecutive frames, rather than at every pixel as in dense optical flow. The Lucas-Kanade method with pyramidal refinement, applied to corners from detectors like Shi-Tomasi, is the classical pipeline. Sparse flow is computationally cheap and underlies feature tracking in visual odometry and visual-inertial SLAM front ends. Why it matters for physical AI: Feature tracking at kilohertz-friendly cost keeps visual odometry running on embedded compute, anchoring ego-motion estimates for drones and mobile manipulators without GPUs. ### Sparse Reward Category: Robot Learning URL: https://almond.bot/glossary/sparse-reward A sparse reward is a reinforcement learning reward signal that is nonzero only at rare events, typically task completion, giving the agent no gradient of progress during the vast majority of exploration. Sparse rewards are easy to specify correctly but make credit assignment and exploration hard, especially over long horizons. Standard remedies include reward shaping, hindsight experience replay, curriculum learning, and seeding with demonstrations. Why it matters for physical AI: Task success is trivial to define but brutal to learn from; managing the specification-versus-learnability trade-off is a recurring theme in training manipulation policies with RL. ### Spatial Attention Category: Perception URL: https://almond.bot/glossary/spatial-attention Spatial attention is a neural network mechanism that weights features by image or spatial location, letting a model focus computation on task-relevant regions of its visual input. In transformer architectures, self- and cross-attention over patch tokens implement it natively; robotic policy networks also use learned keypoint or attention maps over camera views to localize objects and grasp points. Attention maps double as an interpretability signal showing where a policy is looking. Why it matters for physical AI: Manipulation success often hinges on a few pixels around the gripper and target; architectures that attend spatially generalize better across cluttered scenes and camera perturbations. ### Spatial Reasoning Category: Foundation Models URL: https://almond.bot/glossary/spatial-reasoning Spatial reasoning is the capacity of a model to understand and manipulate geometric and relational information: relative positions, distances, orientations, containment, and viewpoint changes. Off-the-shelf vision-language models are notably weak at metric spatial tasks, motivating targeted efforts such as SpatialVLM and spatial question-answering benchmarks that train or probe for 3D-aware understanding. Robotics-oriented models increasingly inject depth, camera geometry, or 3D tokens to compensate. Why it matters for physical AI: A model that cannot judge whether the gripper is left of the handle cannot manipulate reliably; closing the spatial gap in VLM backbones is a live bottleneck for vision-language-action systems. ### Speed and Separation Monitoring Category: Safety & Standards URL: https://almond.bot/glossary/speed-and-separation-monitoring Speed and separation monitoring is a collaborative operation mode defined in ISO/TS 15066 and ISO 10218 in which the robot continuously maintains at least a protective separation distance from humans, computed from robot and human speeds, stopping distances, and sensing latency. Safety-rated sensors such as laser scanners or vision systems track people in the cell, and the robot scales its speed down (ultimately to a stop) as separation shrinks, resuming when clearance is restored. Why it matters for physical AI: Dynamic speed scaling lets robots keep working productively near people instead of halting behind fences, and it exemplifies the certified runtime monitoring that learned controllers must operate within. ### SPI Protocol Category: Hardware URL: https://almond.bot/glossary/spi-protocol SPI (Serial Peripheral Interface) is a synchronous, full-duplex serial communication protocol in which a controller device exchanges data with peripherals over shared clock (SCLK), data-out (MOSI), and data-in (MISO) lines, selecting targets with per-device chip-select signals. It supports clock rates from megahertz to tens of megahertz with minimal protocol overhead, at the cost of no built-in addressing or acknowledgment. Robots use SPI to read IMUs, magnetic encoders, ADCs, and other sensors at high rates. Why it matters for physical AI: Kilohertz control loops need sensor reads with microsecond determinism; SPI's simplicity and speed make it the default link between microcontrollers and the encoders and IMUs closing those loops. ### Spline Interpolation Category: Math & Kinematics URL: https://almond.bot/glossary/spline-interpolation Spline interpolation is the construction of smooth piecewise-polynomial curves passing through or near a sequence of waypoints, widely used to turn sparse via-points into continuous robot trajectories. Cubic splines guarantee continuous velocity and acceleration; quintic splines additionally control jerk, and B-splines offer local control without exact interpolation. Trajectory generators fit splines in joint or Cartesian space and time-parameterize them against velocity and acceleration limits. Why it matters for physical AI: Discontinuous acceleration excites vibration and tracking error on real arms, so spline smoothing sits between virtually every planner or policy and the servo loops executing its output. ### Spot Arm Category: Hardware URL: https://almond.bot/glossary/spot-arm The Spot Arm is Boston Dynamics' six-degree-of-freedom manipulator with an integrated gripper, mounted on the Spot quadruped to form a legged mobile manipulation platform. The arm supports door opening, valve turning, object pickup, and dragging, with whole-body control coordinating arm and legs, and can be teleoperated or scripted through Spot's SDK. It brought mobile manipulation to a commercially deployed legged robot in inspection and industrial settings. Why it matters for physical AI: Legged bases extend manipulation to stairs and rough terrain that wheeled platforms cannot reach, and commercially fielded systems like this supply real-world lessons in whole-body autonomy. ### Spray Coating Robot Category: Applications URL: https://almond.bot/glossary/spray-coating-robot A spray coating robot is an industrial manipulator that applies paint, sealant, or other coatings along programmed paths, maintaining consistent standoff distance, orientation, and speed to achieve uniform film thickness. Paint robots are typically explosion-proof for solvent atmospheres and are programmed by offline path planning over part CAD models or by demonstration. Automotive body painting has been one of the most thoroughly automated robotic applications since the 1980s. Why it matters for physical AI: Coating exemplifies surface-following trajectory generation with tight process tolerances: a template for how offline programming and CAD-driven path planning industrialize repetitive skilled work. ### Stance Phase Category: Locomotion URL: https://almond.bot/glossary/stance-phase The stance phase is the portion of a legged locomotion gait cycle during which a foot is in contact with the ground, supporting load and generating propulsive and stabilizing ground reaction forces. In human walking it spans roughly 60 percent of the cycle, and in robot gait design the stance-to-cycle ratio (duty factor) is a key parameter distinguishing walking from running gaits. Stance-leg control governs balance, force distribution, and odometry-relevant contact estimation. Why it matters for physical AI: Everything a legged robot does to stay upright happens through stance feet; contact scheduling and stance force control are where locomotion policies succeed or fall over. ### State Abstraction Category: Robot Learning URL: https://almond.bot/glossary/state-abstraction State abstraction is the mapping of raw, high-dimensional states or observations into compact representations that preserve the information needed for decision making while discarding irrelevant detail. Formal treatments include bisimulation metrics, which group states with equivalent reward and transition behavior, while practical robot learning realizes abstraction through learned latent states, object-centric representations, or hand-chosen features. Good abstractions shrink the effective problem and improve generalization. Why it matters for physical AI: Cameras deliver megapixels while tasks depend on a handful of object poses; how a policy compresses that gap largely determines its sample efficiency and transfer across scenes. ### State Estimation Category: Control URL: https://almond.bot/glossary/state-estimation State estimation is the inference of a system's internal state (pose, velocity, contact status, or joint configuration) from noisy, partial sensor measurements and a dynamics model. The Kalman filter and its extended and unscented variants handle Gaussian settings, particle filters address multimodal posteriors, and factor graph smoothers fuse measurements over trajectories. Legged robots fuse IMU, kinematics, and contact events to estimate base motion; manipulators estimate object pose and contact forces. Why it matters for physical AI: Controllers and policies act on estimates, not truth; estimator bias or lag propagates directly into instability and failed grasps, making estimation quality a hard floor on performance. ### State Space Category: Control URL: https://almond.bot/glossary/state-space A state space is the set of all possible states of a dynamical system, where a state is a variable collection sufficient to determine future evolution given future inputs. State-space control models dynamics as x' = f(x, u), linearized to x' = Ax + Bu for classical design, while in reinforcement learning the state space is one component of the Markov decision process. Robot state spaces are typically continuous and high-dimensional, combining configurations, velocities, and task variables. Why it matters for physical AI: The size and structure of the state space dictate which methods apply at all (tabular RL, linear control, or deep function approximation) and where the curse of dimensionality bites. ### State-Action Pair Category: Robot Learning URL: https://almond.bot/glossary/state-action-pair A state-action pair is the tuple (s, a) coupling an environment state or observation with the action taken there, the atomic unit of experience in reinforcement and imitation learning. Demonstrations decompose into sequences of state-action pairs for behavior cloning, replay buffers store them within transition tuples (s, a, r, s'), and Q-functions assign values to them. Dataset scale in robot learning is often quoted in state-action pairs or transitions. Why it matters for physical AI: Robot datasets, from teleoperated demos to fleet logs, are ultimately piles of state-action pairs; their quantity, diversity, and labeling quality bound what any policy can learn. ### Static Gait Category: Locomotion URL: https://almond.bot/glossary/static-gait A static gait is a legged locomotion pattern that maintains static stability throughout: the robot's center of mass projection stays inside the support polygon formed by the feet in contact at every instant, so motion can be paused at any point without falling. Quadruped crawl gaits, which move one leg at a time while three remain planted, are the canonical example. Static gaits are slow but robust, whereas dynamic gaits accept transient instability for speed. Why it matters for physical AI: Statically stable walking remains the safe fallback for legged platforms carrying payloads or traversing uncertain terrain, and the static-dynamic distinction frames the entire design space of gait controllers. ### Steady-State Error Category: Control URL: https://almond.bot/glossary/steady-state-error Steady-state error is the residual difference between a control system's output and its reference as time goes to infinity, after transients have decayed. Its magnitude depends on system type and input class: a type-1 system has zero steady-state error to a step but finite error to a ramp, and integral action in PID control is the standard mechanism for eliminating constant offsets. Persistent disturbances such as gravity loading produce steady-state error unless compensated. Why it matters for physical AI: A gripper that settles two millimeters from its target misses the grasp; integral compensation and gravity feedforward exist precisely to drive this error to zero on real arms. ### Step Frequency Category: Locomotion URL: https://almond.bot/glossary/step-frequency Step frequency, or cadence, is the number of steps a legged robot takes per unit time, which together with step length determines locomotion speed. Gait controllers treat it as a design or optimization variable trading off speed, stability, and energy: higher frequencies allow faster reaction to perturbations and smaller individual steps, but raise actuator bandwidth demands and impact losses. Model predictive controllers and learned locomotion policies often adapt cadence to terrain and commanded velocity. Why it matters for physical AI: Cadence selection couples directly to actuator limits and recovery ability; adaptive step timing is one of the mechanisms behind the robustness of modern learned quadruped controllers. ### Stepper Motor Category: Hardware URL: https://almond.bot/glossary/stepper-motor A stepper motor is an electric motor that rotates in discrete angular increments (commonly 200 full steps per revolution) by sequentially energizing stator windings, enabling open-loop position control without feedback sensors. Microstepping drivers subdivide steps for smoother motion and finer resolution. Steppers are inexpensive and hold position when stationary, but can silently lose steps under excess load, which is why 3D printers and low-cost robots use them while dynamic manipulators favor closed-loop servos. Why it matters for physical AI: Cost-sensitive robot designs still lean on steppers for axes with predictable loads, and knowing their lost-step failure mode explains many calibration and drift issues in cheap hardware. ### Stepping Stone Navigation Category: Locomotion URL: https://almond.bot/glossary/stepping-stone-navigation Stepping stone navigation is legged locomotion across terrain offering only sparse, discrete viable footholds (gaps, rubble, or literal stepping stones) where foot placement must be planned precisely rather than chosen freely. It couples perception (identifying safe foothold regions from depth or elevation maps) with footstep planning and dynamic balance under restricted contact choices. Perceptive reinforcement learning controllers, such as those demonstrated in ANYmal parkour work, handle such terrains end to end. Why it matters for physical AI: Constrained footholds are the acid test separating blind gait controllers from perceptive locomotion, driving the integration of exteroception into policies that previously relied on proprioception alone. ### Stereo Camera Category: Hardware URL: https://almond.bot/glossary/stereo-camera A stereo camera is a sensor with two imagers separated by a fixed baseline that recovers depth by triangulating pixel correspondences between the views, with depth accuracy proportional to baseline and inversely related to distance. Active stereo variants like the Intel RealSense D400 series project an infrared texture pattern to aid matching on featureless surfaces. Stereo provides passive, sunlight-tolerant depth, unlike time-of-flight sensors, at the cost of correspondence failures on textureless or reflective regions. Why it matters for physical AI: Affordable dense depth from stereo heads feeds grasp planning, obstacle avoidance, and imitation learning datasets, making these sensors near-ubiquitous on manipulation and mobile platforms alike. ### Stereo Rectification Category: Perception URL: https://almond.bot/glossary/stereo-rectification Stereo rectification is the reprojection of a stereo image pair onto a common virtual image plane so that corresponding points lie on the same horizontal scanline, reducing the correspondence search from two dimensions to one. It is computed from the cameras' intrinsic and extrinsic calibration, warping both images so epipolar lines become horizontal and aligned. Rectification is a standard preprocessing step before disparity estimation in stereo matching pipelines. Why it matters for physical AI: Real-time depth on robots depends on scanline-restricted matching; degraded calibration silently breaks rectification, which is a leading cause of depth artifacts in fielded systems. ### Stiffness Control Category: Control URL: https://almond.bot/glossary/stiffness-control Stiffness control is the regulation of a robot's apparent elastic behavior, commanding the relationship between end-effector (or joint) displacement and restoring force so the robot behaves like a programmable spring about an equilibrium pose. It is the static special case of impedance control, which additionally shapes apparent damping and inertia. Stiffness can be rendered actively through torque control or provided physically by variable stiffness actuators. Why it matters for physical AI: Tasks from peg insertion to polishing succeed by being soft in some directions and stiff in others; commanding stiffness, not just position, is fundamental to contact-rich manipulation. ### Stochastic Gradient Descent (SGD) Category: Robot Learning URL: https://almond.bot/glossary/stochastic-gradient-descent Stochastic gradient descent (SGD) is an optimization algorithm that updates model parameters using gradients estimated from small random mini-batches of data rather than the full dataset, making large-scale neural network training tractable. Variants adding momentum and adaptive per-parameter learning rates, most prominently Adam and AdamW, dominate practice. Virtually every learned robot policy, perception model, and world model is trained with some member of this family. Why it matters for physical AI: Optimizer behavior shapes robot learning outcomes concretely: training stability on small demonstration datasets, fine-tuning dynamics of pretrained VLAs, and convergence of RL losses all trace back to SGD mechanics. ### Stochastic Policy Category: Robot Learning URL: https://almond.bot/glossary/stochastic-policy A stochastic policy is a policy that defines a probability distribution over actions for each state and samples from it, rather than returning a single deterministic action. Gaussian policies over continuous torques or end-effector deltas are standard in policy gradient and maximum-entropy RL, where stochasticity drives exploration and smooths optimization; diffusion policies represent richer multimodal action distributions. Deployed systems often act on the distribution's mean or a low-temperature sample. Why it matters for physical AI: Human demonstrations are multimodal (there are many valid ways to grasp a cup) and policies that model action distributions rather than averages avoid committing to invalid compromises. ### Strain Gauge Category: Hardware URL: https://almond.bot/glossary/strain-gauge A strain gauge is a sensor whose electrical resistance changes proportionally with the mechanical strain of the surface it is bonded to, typically read through a Wheatstone bridge circuit to resolve microvolt-scale signals. Arranged on elastic flexure structures, strain gauges form load cells and six-axis force-torque sensors; silicon strain-sensing variants achieve higher sensitivity. They are the dominant transduction principle behind force sensing in robot wrists, fingers, and joints. Why it matters for physical AI: Force-aware manipulation (insertion, polishing, handovers) bottoms out in strain measurement; gauge quality, temperature drift, and calibration set the fidelity of every force control loop above them. ### Strain Wave Gear Category: Hardware URL: https://almond.bot/glossary/strain-wave-gear A strain wave gear, widely known by the trade name harmonic drive, is a compact speed reducer consisting of an elliptical wave generator, a flexible externally toothed flexspline, and a rigid internally toothed circular spline, achieving reduction ratios of roughly 50:1 to 160:1 in a single stage. Because many teeth engage simultaneously, it offers near-zero backlash, high torque density, and compact form. It is the standard joint reducer in industrial and collaborative robot arms. Why it matters for physical AI: Joint transmissions define an arm's precision and back-drivability; the strain wave gear's zero-backlash accuracy (and its compliance and friction quirks) pervade manipulator control design. ### Stretch (Hello Robot) Category: Hardware URL: https://almond.bot/glossary/stretch Stretch is a lightweight mobile manipulator from Hello Robot, combining a differential-drive base, a vertical lift, a horizontally telescoping arm, and a compliant wrist and gripper in a slim, human-friendly footprint. Its unusual prismatic kinematics keep cost, weight, and inertia low relative to arm-on-base designs, and Stretch 3 ships with an RGB-D head camera and open-source ROS 2 software. It is widely used for in-home assistive robotics and mobile manipulation research. Why it matters for physical AI: Homes demand safe, affordable, and narrow robots more than dexterous ones; minimalist platforms like this test how much everyday utility careful hardware-software co-design can unlock. ### Structure from Motion Category: Perception URL: https://almond.bot/glossary/structure-from-motion Structure from motion is the reconstruction of 3D scene geometry and camera poses from a collection of overlapping 2D images, by detecting and matching features across views, estimating relative camera geometry, triangulating points, and refining everything with bundle adjustment. Pipelines like COLMAP produce sparse point clouds and calibrated poses, often seeding dense multi-view stereo or radiance field reconstruction. Unlike SLAM, it typically runs offline over unordered image sets. Why it matters for physical AI: Camera-only 3D reconstruction turns commodity images into maps, object models, and simulation assets, including the real-to-sim scene capture pipelines behind photorealistic robot training environments. ### Structured Prediction Category: Robot Learning URL: https://almond.bot/glossary/structured-prediction Structured prediction is machine learning that outputs structured objects (sequences, trees, graphs, or segmentations) whose components have mutual dependencies, rather than independent scalar labels. Classical formulations include conditional random fields and structured SVMs; in modern practice, autoregressive decoding and diffusion serve the same role. Robotics outputs are inherently structured: action sequences, grasp poses on SE(3), scene graphs, and trajectories all demand coherent joint prediction rather than independent guesses. Why it matters for physical AI: Action chunking, trajectory generation, and scene graph inference are structured prediction problems in disguise; recognizing this explains why sequence models and diffusion dominate modern policy architectures. ### Subgoal Category: Robot Learning URL: https://almond.bot/glossary/subgoal A subgoal is an intermediate target state or condition that decomposes a long-horizon task into shorter, independently achievable segments, such as "grasp the handle" within "open the drawer and retrieve the tool". Hierarchical methods have a high-level process (a learned high-level policy, a symbolic planner, or a language model) propose subgoals that a low-level goal-conditioned policy then reaches; subgoals may be expressed as states, images, language, or latent vectors. Why it matters for physical AI: Credit assignment across thousands of timesteps defeats flat policies; subgoal decomposition is the main lever for extending today's short-horizon skills to multi-stage real-world tasks. ### Success Rate Category: Data & Benchmarks URL: https://almond.bot/glossary/success-rate Success rate is the fraction of evaluation trials in which a robot completes a task according to a predefined success criterion, and it is the primary headline metric in manipulation and robot learning research. Meaningful reporting requires specifying the number of trials, initial condition distribution, object sets, and success detection method, since small trial counts carry wide confidence intervals. Cross-lab comparability remains poor because physical setups differ in unmeasured ways. Why it matters for physical AI: Claims about robot foundation models rise or fall on evaluation rigor; understanding what a reported success rate does and does not control for is essential to interpreting progress. ### Successor Features Category: Robot Learning URL: https://almond.bot/glossary/successor-features Successor features are a reinforcement learning representation (Barreto et al., 2017) that expresses a policy's value as the dot product between the expected discounted sum of future state features and a reward weight vector, decoupling environment dynamics from task reward. They generalize Dayan's successor representation (1993) and enable rapid transfer: when the reward changes but dynamics do not, new values follow from re-weighting cached successor features, supporting generalized policy improvement across task families. Why it matters for physical AI: Robots amortize hardware cost by performing many tasks in one environment; representations that reuse dynamics knowledge across reward specifications directly serve that multi-task economics. ### Suction Grasping Category: Manipulation URL: https://almond.bot/glossary/suction-grasping Suction grasping is object acquisition using vacuum pressure through one or more compliant suction cups, requiring only a single reachable, sufficiently flat and non-porous surface patch rather than antipodal contact geometry. It dominates logistics picking for boxes, polybags, and consumer goods, and featured centrally in Amazon Picking Challenge entries. Dex-Net 3.0 formalized learned suction grasp quality prediction, modeling seal formation and wrench resistance from depth images. Why it matters for physical AI: A single-point grasp with millimeter pose tolerance is vastly easier to automate at scale, which is why suction underpins most deployed warehouse picking and mixed-bin handling systems. ### Supervised Fine-Tuning (SFT) Category: Foundation Models URL: https://almond.bot/glossary/supervised-fine-tuning Supervised fine-tuning (SFT) is the adaptation of a pretrained model by continued training on curated input-output pairs with a standard supervised loss, specializing broad pretrained capabilities to a target behavior. In robot foundation models, training a vision-language-action model on teleoperated demonstrations is effectively SFT of a vision-language backbone with action outputs. It typically precedes further post-training stages such as reinforcement learning or preference optimization. Why it matters for physical AI: The pretrain-then-SFT recipe imported from language modeling is now the dominant way robot policies inherit internet-scale visual and semantic knowledge before learning to act. ### Support Vector Machine Category: Robot Learning URL: https://almond.bot/glossary/support-vector-machine A support vector machine is a supervised learning algorithm that finds the maximum-margin hyperplane separating classes, using the kernel trick to obtain nonlinear decision boundaries in implicit high-dimensional feature spaces (Cortes and Vapnik, 1995). Before deep learning, SVMs over hand-crafted features were standard in robotics for grasp stability classification, terrain and traversability classification, and object recognition. They remain useful for small-data problems with engineered features. Why it matters for physical AI: Small labeled datasets are common at the edges of robotics (fault detection, contact classification) where margin-based methods on good features still beat data-hungry networks. ### Surgical Robot Category: Applications URL: https://almond.bot/glossary/surgical-robot A surgical robot is a medical robotic system that assists or performs surgical procedures, exemplified by Intuitive Surgical's da Vinci, a teleoperated platform for minimally invasive surgery providing tremor filtering, motion scaling, and wristed instruments under surgeon control. Research systems push toward autonomy (the STAR robot demonstrated supervised autonomous intestinal anastomosis) and the open da Vinci Research Kit (dVRK) supports academic work on surgical automation and shared control. Why it matters for physical AI: Surgery sets the ceiling for precision, latency, and safety validation in teleoperation and shared autonomy, and its regulatory pathway previews how high-stakes robot autonomy gets certified. ### Swarm Robotics Category: Applications URL: https://almond.bot/glossary/swarm-robotics Swarm robotics is the coordination of large numbers of relatively simple robots through decentralized, local interactions, drawing on collective behaviors of social insects such as stigmergy and flocking. Desired global behavior (aggregation, formation, collective transport, area coverage) emerges without central control, conferring scalability and fault tolerance. Harvard's Kilobot project, which demonstrated programmable self-assembly with 1,024 robots, is a landmark experiment. Why it matters for physical AI: Fleet-scale deployment raises swarm questions even for sophisticated robots: decentralized coordination, communication limits, and graceful degradation matter whenever many machines share a warehouse or field. ### Swing Phase Category: Locomotion URL: https://almond.bot/glossary/swing-phase The swing phase is the portion of a legged gait cycle during which a foot is off the ground, traveling from liftoff to its next touchdown location. Swing-leg control shapes the foot trajectory for ground clearance, places the foot to regulate balance (following principles like Raibert's foot placement heuristic) and manages touchdown timing and impact. Insufficient swing clearance causes tripping, a dominant failure mode on rough terrain. Why it matters for physical AI: Foot placement during swing is the primary control authority a legged robot has over its own stability, which is why learned locomotion policies devote much of their capacity to it. ### Symplectic Integration Category: Simulation URL: https://almond.bot/glossary/symplectic-integration Symplectic integration is a class of numerical methods for Hamiltonian systems that exactly preserve the symplectic structure of phase space, yielding bounded long-term energy error instead of the systematic drift produced by non-symplectic schemes like explicit Euler. The semi-implicit (symplectic) Euler method, which updates velocity before position, is the workhorse of physics engines including MuJoCo's default integrator. Higher-order members include the Verlet and leapfrog families. Why it matters for physical AI: Simulated training runs billions of timesteps, so integrators that stay stable at large steps without energy explosion directly determine simulation throughput and the physical plausibility of learned behaviors. ### System Identification Category: Control URL: https://almond.bot/glossary/system-identification System identification is the construction of mathematical models of a dynamical system from measured input-output data, estimating parameters such as link masses, inertias, friction coefficients, and actuator characteristics, or fitting non-parametric and learned dynamics models. In robotics it calibrates model-based controllers and narrows the sim-to-real gap by tuning simulator parameters to match hardware (the real-to-sim direction) with excitation trajectories designed to make parameters observable. Why it matters for physical AI: Accurate dynamics models multiply the value of simulation and model-based control alike; identification is frequently the cheapest single step toward transferring a simulation-trained policy onto hardware. ## Glossary: T ### Tabletop Manipulation Category: Manipulation URL: https://almond.bot/glossary/tabletop-manipulation Tabletop manipulation is a canonical robotics setting in which a robot arm interacts with objects on a flat, bounded work surface, typically performing pick-and-place, stacking, sorting, or rearrangement tasks. Its constrained geometry, easy resettability, and compatibility with fixed overhead or wrist-mounted cameras have made it the dominant testbed for manipulation research, from classical grasp planning to benchmarks like Ravens and many of the demonstrations in Open X-Embodiment. Why it matters for physical AI: Most large-scale robot manipulation datasets and policy evaluations are collected in tabletop scenes, so the setting effectively defines what current robot foundation models are good at, and the gap between tabletop success and open-world performance is a central deployment challenge. ### Tactile Sensing Category: Perception URL: https://almond.bot/glossary/tactile-sensing Tactile sensing is the measurement of physical contact between a robot and its environment, including contact location, normal and shear forces, vibration, and texture. Sensor technologies span resistive and capacitive arrays, barometric taxels, and vision-based sensors such as GelSight and DIGIT, which image the deformation of an illuminated elastomer to recover high-resolution contact geometry. Tactile signals complement vision precisely where cameras fail: under occlusion by the gripper, during in-hand manipulation, and for perceiving properties like friction, slip, and compliance. Why it matters for physical AI: Contact-rich tasks such as insertion, cable routing, and in-hand reorientation are hard to solve from vision alone, and incorporating touch into learned policies and foundation-model architectures is an active frontier for making manipulation robust in the real world. ### Tactile Skin Category: Hardware URL: https://almond.bot/glossary/tactile-skin Tactile skin is a flexible, large-area array of touch sensors designed to cover curved robot surfaces such as arms, torsos, or fingers, providing distributed contact sensing rather than measurement at a single fingertip. Implementations use capacitive, resistive, piezoelectric, or magnetic taxel grids, often with serialized readout to keep wiring tractable. Whole-body skins enable robots to detect and localize incidental contact anywhere on their structure, not just at the end-effector. Why it matters for physical AI: Distributed contact awareness underpins safe operation around people and enables whole-arm manipulation strategies, and large-area touch is increasingly viewed as a missing data modality for training more capable manipulation foundation models. ### Task and Motion Planning (TAMP) Category: Manipulation URL: https://almond.bot/glossary/task-and-motion-planning Task and motion planning (TAMP) is the integrated problem of choosing a discrete sequence of symbolic actions, such as pick, place, or open, together with the continuous motions, grasps, and placements that realize each action. Because the feasibility of a symbolic plan depends on geometric details like reachability and collision-free paths, TAMP solvers interleave symbolic search with continuous constraint satisfaction, as in systems like PDDLStream. It remains the standard formulation for long-horizon manipulation with strong correctness guarantees. Why it matters for physical AI: Long-horizon tasks still challenge end-to-end learned policies, and hybrid systems that pair foundation-model task reasoning with TAMP-style geometric verification are a leading approach to reliable multi-step manipulation in deployment. ### Task Space Category: Math & Kinematics URL: https://almond.bot/glossary/task-space Task space is the coordinate space in which a robot's goal is naturally expressed, most commonly the 6-DoF Cartesian pose of the end-effector, as opposed to joint space, which describes the robot's internal configuration. Controllers and planners map between the two using the forward kinematics and the Jacobian, and task-space control laws such as operational space control command end-effector wrenches or velocities directly. Redundant robots can exploit extra degrees of freedom in the null space of the task. Why it matters for physical AI: Learned manipulation policies overwhelmingly output task-space end-effector actions because they transfer more easily across robots with different kinematics, a convention that underlies most cross-embodiment datasets and robot foundation models. ### Task Specification Category: Robot Learning URL: https://almond.bot/glossary/task-specification Task specification is the way a desired behavior is communicated to a robot, whether as a reward function, a goal state or goal image, a natural-language instruction, a demonstration, or a formal logical specification. The choice determines what supervision the system needs and what ambiguities remain: rewards are precise but hard to author, language is flexible but underspecified, and demonstrations convey intent implicitly. Mis-specified tasks lead to reward hacking and unintended behavior. Why it matters for physical AI: Language-conditioned robot foundation models have made natural-language instruction the dominant interface for commanding robots, which makes understanding what a specification does and does not pin down critical for safe, predictable deployment. ### Task Success Rate Category: Data & Benchmarks URL: https://almond.bot/glossary/task-success-rate Task success rate is the fraction of evaluation episodes in which a robot policy achieves a task's success criterion, and it is the primary headline metric in manipulation and locomotion research. Reported rates depend heavily on evaluation protocol: the number of trials, object and scene randomization, distractor placement, initial-state distribution, and how success is judged, whether automatically in simulation or by a human. Small trial counts produce wide confidence intervals that are frequently omitted. Why it matters for physical AI: Comparing robot foundation models rigorously requires standardized, statistically sound success-rate evaluation, and the gap between papers reporting on curated scenes and performance in uncontrolled deployments is a persistent source of overestimated capability. ### Task-Parameterized Learning Category: Robot Learning URL: https://almond.bot/glossary/task-parameterized-learning Task-parameterized learning is a family of learning-from-demonstration methods that encode a skill relative to multiple task-relevant coordinate frames, such as frames attached to objects, start poses, and goals, so the learned motion adapts automatically when those frames move. The canonical formulation, task-parameterized Gaussian mixture models developed by Calinon and colleagues, learns local models in each frame and fuses them at reproduction time. This yields strong generalization from few demonstrations without retraining. Why it matters for physical AI: Adapting demonstrated skills to new object poses with a handful of examples remains valuable for data-efficient deployment, and the underlying idea of frame-relative encoding echoes in modern object-centric action representations for learned policies. ### TD3 (Twin Delayed DDPG) Category: Robot Learning URL: https://almond.bot/glossary/td3 TD3 (Twin Delayed Deep Deterministic Policy Gradient) is an off-policy actor-critic reinforcement learning algorithm for continuous control, introduced by Fujimoto et al. in 2018 as a successor to DDPG. It curbs the overestimation bias that destabilizes DDPG through three mechanisms: twin critics with a minimum-value target, delayed policy updates, and target policy smoothing via clipped noise. TD3 remains a standard baseline alongside SAC and PPO on continuous-control benchmarks. Why it matters for physical AI: Sample-efficient off-policy algorithms like TD3 make reinforcement learning tractable for robot control, where every real-world trial is expensive, and its stabilization tricks inform the training of value functions used in modern robot learning pipelines. ### Teach Pendant Category: Hardware URL: https://almond.bot/glossary/teach-pendant A teach pendant is a handheld control terminal connected to an industrial robot controller, used to jog the robot through waypoints, record positions, edit programs, and manage operating modes. Pendants typically include a touchscreen or keypad, a three-position enabling switch that must be held in its middle position for motion, and an emergency stop, as required by industrial robot safety standards. Waypoint programming via pendant remains the dominant method for deploying traditional industrial arms. Why it matters for physical AI: The slow, expert-driven pendant workflow is precisely what learned policies and demonstration-based programming aim to replace, making it the baseline against which the economics of foundation-model-driven automation are judged. ### Teacher-Student Distillation Category: Robot Learning URL: https://almond.bot/glossary/teacher-student-distillation Teacher-student distillation is a training paradigm in which a teacher policy with access to privileged information, such as ground-truth terrain geometry, contact states, or object poses available only in simulation, supervises a student policy that must act from deployable observations like proprioception and cameras. The approach was popularized in legged locomotion by Lee et al.'s 2020 work on the ANYmal quadruped and is now standard in sim-to-real reinforcement learning pipelines. Why it matters for physical AI: Distillation lets reinforcement learning exploit simulator-only state during training while producing policies that run on real sensors, and the same recipe is used to compress large robot foundation models into fast onboard controllers. ### TEB Planner (Timed Elastic Band) Category: Navigation & SLAM URL: https://almond.bot/glossary/teb-planner The TEB (Timed Elastic Band) planner is a local trajectory planner for mobile robots that deforms an initial path into a time-parameterized trajectory by optimizing waypoint poses and time intervals jointly, subject to obstacle clearance, velocity and acceleration limits, and kinodynamic constraints such as car-like steering. Introduced by Rösmann et al. and implemented as a ROS local planner, it optimizes several topologically distinct candidate trajectories in parallel to avoid poor local minima. Why it matters for physical AI: Reactive local planners like TEB remain the workhorse layer that turns global routes into safe, dynamically feasible motion for warehouse and service robots, sitting beneath whatever learned or foundation-model navigation stack issues the high-level goals. ### Telemetry Category: Software & Middleware URL: https://almond.bot/glossary/telemetry Telemetry is the automated collection and transmission of a robot's operational data, including joint states, sensor readings, battery status, controller health, error events, and task outcomes, to remote monitoring and logging infrastructure. Robot telemetry systems typically stream time-series data over protocols such as MQTT, DDS, or gRPC into fleet dashboards and long-term stores, with bandwidth-aware handling of high-rate signals like video and point clouds. Why it matters for physical AI: Fleet telemetry is the raw material of the deployment data flywheel: logged episodes surface failure modes, feed incident triage, and become training data that improves robot foundation models over successive versions. ### Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/teleoperation Teleoperation is the control of a robot by a human operator who is physically separated from it, with operator commands mapped to robot motion through interfaces ranging from joysticks and VR controllers to kinematically matched leader arms and whole-body exoskeletons. Classical bilateral teleoperation adds force feedback so the operator feels remote contact. Low-cost leader-follower systems such as ALOHA and handheld devices like UMI have made teleoperation the dominant method for collecting robot manipulation demonstrations. Why it matters for physical AI: Nearly all high-quality manipulation training data behind current robot foundation models originates from teleoperated demonstrations, so teleoperation throughput, ergonomics, and fidelity directly bound how fast the field can scale imitation learning. ### Temporal Abstraction Category: Robot Learning URL: https://almond.bot/glossary/temporal-abstraction Temporal abstraction is the structuring of sequential decision-making at multiple time scales, so that a high-level policy selects extended behaviors, such as options, skills, or subgoals, while low-level policies handle fine-grained control within each behavior. The options framework of Sutton, Precup, and Singh formalized this in reinforcement learning, and hierarchical policies, skill discovery, and action chunking are all instances. Abstraction shortens effective horizons and simplifies credit assignment. Why it matters for physical AI: Long-horizon household and industrial tasks are far easier to learn and plan over when decomposed into reusable skills, and modern systems pair vision-language models issuing abstract subgoals with low-level policies executing them. ### Temporal Difference Learning Category: Robot Learning URL: https://almond.bot/glossary/temporal-difference-learning Temporal difference (TD) learning is a class of reinforcement learning methods that update value estimates from the difference between successive predictions, bootstrapping from the current estimate of future return rather than waiting for complete episodes. Introduced by Sutton in 1988, TD learning underlies Q-learning, SARSA, and the critic updates in actor-critic algorithms, and the TD error itself parallels reward-prediction-error signals observed in dopamine neurons. Why it matters for physical AI: Bootstrapped value estimation is what makes reinforcement learning sample-efficient enough to consider for real robots, and TD-based critics remain core components in the RL fine-tuning stages of modern robot policy training. ### Temporal Ensemble Category: Robot Learning URL: https://almond.bot/glossary/temporal-ensemble Temporal ensembling is an inference-time technique for action-chunking policies in which overlapping action chunks predicted at successive time steps are averaged, typically with exponential weighting, to produce the action actually executed. Introduced with the ACT policy in the ALOHA system by Zhao et al. in 2023, it smooths discontinuities at chunk boundaries and reduces the impact of any single erroneous prediction without retraining the policy. Why it matters for physical AI: Simple inference-time smoothing measurably improves the precision and reliability of imitation-learned manipulation policies, and temporal ensembling has become a default component in the action-chunking recipes used across modern robot foundation models. ### Tendon-Driven Actuation Category: Hardware URL: https://almond.bot/glossary/tendon-driven-actuation Tendon-driven actuation is a transmission scheme in which motors move a robot's joints remotely through cables or tendons routed over pulleys and through sheaths, allowing actuators to be placed away from the joints they drive. This keeps distal links light, which is why tendon drives dominate anthropomorphic hand designs, including the Shadow Hand, and appear in some full arms and surgical robots. Drawbacks include friction, elasticity, creep, and the need for tension maintenance and periodic recalibration. Why it matters for physical AI: Lightweight, human-proportioned hands and arms for humanoids depend on remote actuation, and the compliance and friction of tendon transmissions shape both the control problem and the sim-to-real gap for dexterous manipulation policies. ### Tensor Decomposition Category: Math & Kinematics URL: https://almond.bot/glossary/tensor-decomposition Tensor decomposition is the factorization of a multi-dimensional array into structured combinations of simpler components, with canonical forms including CP decomposition, Tucker decomposition, and tensor-train formats. In robotics and machine learning it is used to compress neural network weights, expose low-rank structure in multi-modal sensor data, and represent value functions or dynamics models over structured state spaces compactly. Why it matters for physical AI: Compressing large policy and perception networks with low-rank factorizations helps fit foundation-model-scale computation onto power-constrained onboard hardware, an increasingly important lever as robot policies grow toward billions of parameters. ### Terminal Sliding Mode Control Category: Control URL: https://almond.bot/glossary/terminal-sliding-mode-control Terminal sliding mode control is a variant of sliding mode control whose sliding surface includes fractional-power terms, driving tracking errors to zero in finite time rather than merely asymptotically. Like all sliding mode methods it offers strong robustness to matched disturbances and model uncertainty, and nonsingular formulations avoid the unbounded control actions of early designs. Chattering from the discontinuous switching term remains the main practical drawback, mitigated by boundary layers or higher-order sliding modes. Why it matters for physical AI: Finite-time convergent, disturbance-robust controllers are attractive low-level layers beneath learned policies on manipulators and quadrotors, guaranteeing tracking of commanded trajectories even when payloads and contacts vary unpredictably. ### Terrain Estimation Category: Perception URL: https://almond.bot/glossary/terrain-estimation Terrain estimation is the online reconstruction of the geometry and physical properties of the ground around a mobile robot, typically as elevation maps, traversability maps, or friction and compliance estimates built from depth cameras, LiDAR, and proprioception. Robot-centric elevation mapping frameworks fuse range measurements with pose estimates while handling drift and occlusion, and learned methods infer terrain properties that vision alone cannot observe, such as slipperiness or deformability, from contact feedback. Why it matters for physical AI: Legged and wheeled robots operating off curated floors must anticipate what the ground will do under load, and terrain-aware perception is a key ingredient separating lab locomotion demos from field deployment. ### Terrain Generation Category: Simulation URL: https://almond.bot/glossary/terrain-generation Terrain generation is the procedural creation of simulated ground geometry, such as slopes, stairs, steps, gaps, rough noise fields, and stepping stones, used to train and evaluate locomotion policies. Modern reinforcement learning pipelines pair generated terrains with curricula that escalate difficulty as the policy improves, a recipe central to massively parallel training in simulators like Isaac Gym. Terrain diversity in simulation is a primary driver of real-world robustness. Why it matters for physical AI: Locomotion policies generalize roughly to the distribution of terrain they trained on, so scalable procedural generation directly determines whether quadrupeds and humanoids can handle stairs, rubble, and slopes outside the lab. ### Tesla Optimus Category: Hardware URL: https://almond.bot/glossary/tesla-optimus Tesla Optimus is a general-purpose bipedal humanoid robot under development by Tesla since 2021, intended for repetitive industrial and eventually household tasks. The platform features Tesla-designed actuators, multi-fingered hands, and a vision-centric autonomy stack that leverages the company's automotive AI infrastructure, with training that draws on teleoperated demonstrations and human video. Successive prototypes have demonstrated walking, object sorting, and manipulation, though the system remains pre-commercial. Why it matters for physical AI: Optimus exemplifies the bet that automotive-scale manufacturing, vertically integrated hardware, and large-scale learned autonomy can make general-purpose humanoids economically viable, shaping investment and talent flows across the physical AI industry. ### Test-Time Adaptation Category: Robot Learning URL: https://almond.bot/glossary/test-time-adaptation Test-time adaptation is the adjustment of a trained model during deployment, without access to original training labels, to cope with distribution shift in the inputs it encounters. Techniques include entropy minimization on unlabeled test data, updating normalization statistics, self-supervised auxiliary objectives, and, in robotics, online system identification or rapid adaptation modules that infer environment parameters from recent experience, as in the RMA approach for legged locomotion. Why it matters for physical AI: Real environments drift away from any training distribution through lighting, wear, payload, and terrain changes, and policies that adapt online rather than failing silently are substantially more viable for long-duration autonomous deployment. ### TF (Transforms) Category: Software & Middleware URL: https://almond.bot/glossary/tf-transforms TF (Transforms) is the ROS subsystem that tracks the time-varying rigid-body transforms among a robot's coordinate frames, such as map, odom, base_link, and each sensor and link frame, organized as a tree. Nodes broadcast transforms and query the buffered tree to express data from any frame in any other at a requested timestamp, with interpolation between samples. The current implementation, tf2, is foundational to virtually every ROS-based perception and navigation stack. Why it matters for physical AI: Consistent, time-aligned coordinate frames are prerequisites for fusing cameras, LiDAR, and kinematics into the observations that learned policies consume, and TF errors remain one of the most common integration failure modes in robot deployments. ### Thermal Camera Category: Perception URL: https://almond.bot/glossary/thermal-camera A thermal camera is an imaging sensor that measures long-wave infrared radiation emitted by objects, producing images of surface temperature rather than reflected light. Uncooled microbolometer arrays are the common variant on robots, operating in the 8-14 micrometer band and functioning in complete darkness, smoke, and fog. Robotics applications include human detection for safety, inspection of electrical and mechanical equipment for overheating, firefighting, and search and rescue. Why it matters for physical AI: Thermal imaging gives robots a perception channel that is robust to lighting conditions that defeat RGB cameras, and it provides reliable human detection for safety systems in dark or visually degraded industrial environments. ### Thermal Coefficient of Expansion Category: Hardware URL: https://almond.bot/glossary/thermal-coefficient-of-expansion The thermal coefficient of expansion (CTE) is a material property quantifying the fractional change in a component's dimensions per degree of temperature change, typically expressed in parts per million per kelvin. In robot design, CTE mismatch between joined materials, such as aluminum structures and carbon fiber or steel, induces stress and geometric distortion as temperature varies, degrading kinematic calibration and repeatability. Precision machines counter this with low-CTE materials, thermal symmetry, and warm-up compensation. Why it matters for physical AI: Robots that hold sub-millimeter accuracy across shifts and seasons need thermally aware mechanical design, since calibration drift from expansion silently erodes the precision that vision-based policies and fixtures assume. ### TidyBot Category: Robot Learning URL: https://almond.bot/glossary/tidybot TidyBot is a 2023 research system from Stanford and Princeton in which a mobile manipulator tidies rooms by putting objects away according to personalized rules, such as which shelf shirts belong on, inferred by a large language model from a handful of user examples. The system combined LLM-based generalization of user preferences with open-vocabulary perception and scripted manipulation primitives, and it became an early landmark for LLM-driven personalization in household robotics. Why it matters for physical AI: Household robots must respect idiosyncratic human preferences rather than a single correct behavior, and TidyBot demonstrated that language models can supply that commonsense personalization layer on top of conventional robot skills. ### Time Synchronization Category: Software & Middleware URL: https://almond.bot/glossary/time-synchronization Time synchronization is the alignment of clocks and data timestamps across a robot's sensors, computers, and actuators so that measurements can be fused coherently. Techniques range from network protocols such as NTP and IEEE 1588 PTP to hardware triggering and pulse-per-second signals that align camera exposures with LiDAR sweeps and IMU samples. Millisecond-level errors are enough to corrupt visual-inertial odometry, sensor fusion, and the observation-action alignment of learned policies. Why it matters for physical AI: Training data for robot foundation models pairs observations with actions at specific instants, so timestamp skew during collection or inference translates directly into degraded policy precision and hard-to-diagnose deployment failures. ### Time-of-Flight Sensor (ToF) Category: Perception URL: https://almond.bot/glossary/time-of-flight-sensor A time-of-flight sensor is a range sensor that measures distance from the round-trip travel time of emitted light, either directly with pulsed lasers or indirectly from the phase shift of modulated illumination. ToF depth cameras capture full depth images at video rates without the texture dependence of stereo, though they contend with multipath interference, ambient sunlight, and reflective or absorptive surfaces. LiDAR is the scanning, long-range member of the same family. Why it matters for physical AI: Compact, low-cost depth sensing feeds the 3D observations that manipulation and navigation policies rely on, and the specific noise characteristics of ToF sensors are an important component of the sim-to-real gap for depth-based policies. ### Tire Handling Robot Category: Applications URL: https://almond.bot/glossary/tire-handling-robot A tire handling robot is an industrial robot system configured to grasp, transport, stack, and palletize tires in tire manufacturing plants, warehouses, and vehicle assembly lines. Tires are challenging payloads: they are heavy, deformable, and often stacked in unstructured piles, so systems combine high-payload arms or gantries, mechanical or vacuum grippers shaped to the bead or tread, and 3D vision for localizing individual tires in bins and stacks. Why it matters for physical AI: Deformable, heavy, semi-structured payloads like tires sit exactly at the boundary where classical fixtured automation struggles, making the category a proving ground for vision-guided grasping and learned bin-picking policies in industrial settings. ### Titanium Category: Hardware URL: https://almond.bot/glossary/titanium Titanium is a structural metal prized in robotics for its high strength-to-weight ratio, roughly half the density of steel at comparable strength, along with corrosion resistance, biocompatibility, and a relatively low thermal expansion coefficient. Its cost and difficult machining restrict it to weight-critical or demanding components, such as legged robot structural parts, surgical robot instruments, prosthetics, and aerospace manipulators, with titanium 3D printing easing some manufacturing constraints. Why it matters for physical AI: Every gram removed from a moving link reduces actuator torque requirements and energy use, so material choices like titanium versus aluminum or carbon fiber shape the payload, endurance, and safety envelope of dynamic robots. ### Tokenization Category: Foundation Models URL: https://almond.bot/glossary/tokenization Tokenization is the conversion of raw data into discrete tokens that a transformer can consume, spanning subword text tokens, image patches, and, in robotics, discretized states and actions. Robot policies tokenize continuous actions by binning each dimension, as in RT-1 and RT-2, or by learning compressed action codes with methods like FAST used in pi0-style models, letting a single sequence model treat perception, language, and control uniformly. Why it matters for physical AI: How actions and observations are tokenized governs the precision, inference speed, and cross-embodiment flexibility of robot foundation models, and action tokenization schemes are an active design axis distinguishing current vision-language-action architectures. ### Tool Changer Category: Hardware URL: https://almond.bot/glossary/tool-changer A tool changer is a coupling mechanism mounted between a robot's wrist flange and its end-effectors that allows tools to be swapped, either automatically under program control or manually with quick-release mechanics. Automatic tool changers lock a master plate to interchangeable tool plates while passing through electrical signals, pneumatics, fluids, and data, letting one arm alternate among grippers, welders, screwdrivers, and sanders within a single cycle. Why it matters for physical AI: Multi-purpose robot cells depend on fast, repeatable tool exchange, and for learned manipulation systems the ability to switch end-effectors extends a single policy platform across tasks that no one gripper can cover. ### Tool Use Category: Manipulation URL: https://almond.bot/glossary/tool-use Tool use is a robot's employment of grasped objects, such as spatulas, screwdrivers, hooks, or hammers, to act on the world in ways its bare end-effector cannot. It requires reasoning about the tool's extended kinematics and contact dynamics, selecting grasps that afford the intended action, and adapting to the tool's geometry, and it has long served as a benchmark of physical intelligence spanning classical affordance learning through modern language-conditioned policies. Why it matters for physical AI: Human environments are built around tools, so general-purpose robots must wield them rather than requiring specialized hardware per task, making tool use a key capability milestone for manipulation foundation models. ### Tool-Augmented Policy Category: Foundation Models URL: https://almond.bot/glossary/tool-augmented-policy A tool-augmented policy is a robot decision-making system, typically built around a large language or vision-language model, that can invoke external software tools during operation, such as perception modules, motion planners, calculators, databases, or code execution, rather than producing all outputs from its own weights. Systems in the Code-as-Policies lineage exemplify the pattern, with the model composing calls to perception APIs and control primitives to solve tasks it was never explicitly trained on. Why it matters for physical AI: Delegating geometry, arithmetic, and safety-critical checks to verified tools compensates for known weaknesses of end-to-end models, and tool-calling architectures offer a pragmatic path to reliable long-horizon autonomy with auditable intermediate steps. ### Torque Control Category: Control URL: https://almond.bot/glossary/torque-control Torque control is a robot control mode in which the commanded quantity at each joint is torque rather than position or velocity, giving direct authority over the forces the robot exerts. It underpins impedance control, gravity compensation, and compliant contact behavior, but demands accurate dynamics models, high-bandwidth current control, and either joint torque sensors or well-characterized actuators. Torque-controlled platforms such as collaborative arms and modern quadrupeds behave gently under unexpected contact. Why it matters for physical AI: Contact-rich manipulation and dynamic legged locomotion both hinge on regulating interaction forces rather than just positions, and torque-level actuation is what allows learned policies to express compliant, safe behavior on real hardware. ### Torque Density Category: Hardware URL: https://almond.bot/glossary/torque-density Torque density is the torque an actuator produces per unit mass or volume, a headline figure of merit for robot joint design. High torque density traditionally comes from high gear ratios, which add friction, backlash, and reflected inertia, whereas quasi-direct-drive designs pair large-diameter, high-pole-count motors with low gear ratios to retain backdrivability and impact robustness, an approach popularized by the MIT Mini Cheetah and now standard in quadrupeds and emerging humanoids. Why it matters for physical AI: Actuator torque density sets the payload, speed, and dynamism achievable at a given robot weight, and progress here is a primary hardware bottleneck for affordable, capable humanoid platforms. ### Torque Ripple Category: Hardware URL: https://almond.bot/glossary/torque-ripple Torque ripple is the periodic variation in an electric motor's output torque as its rotor turns, arising from cogging between magnets and stator slots, non-ideal back-EMF waveforms, and current measurement or commutation errors. In robot joints it produces velocity fluctuations, vibration, and acoustic noise that are most pronounced at low speeds, degrading smooth trajectory tracking and clean force estimation. Mitigations include skewed magnets, optimized windings, and ripple-compensating current control. Why it matters for physical AI: Smooth low-speed motion and clean proprioceptive force sensing both suffer under ripple, which matters for robots that estimate contact from motor currents instead of dedicated force sensors and for precision tasks like polishing and insertion. ### Torque Sensor Category: Hardware URL: https://almond.bot/glossary/torque-sensor A torque sensor is a transducer that measures the torque transmitted through a shaft or joint, most commonly via strain gauges on an elastic element, with capacitive and optical deflection-based designs as alternatives. In robotics, joint torque sensors placed at each axis enable joint-level force control and sensitive collision detection, as in the DLR lightweight robot lineage and KUKA iiwa, while six-axis force-torque sensors at the wrist measure full interaction wrenches at the end-effector. Why it matters for physical AI: Direct torque measurement enables the compliant control and fast contact detection that make robots safe alongside people, and interaction force data enriches demonstrations used to train contact-aware manipulation policies. ### TossingBot Category: Manipulation URL: https://almond.bot/glossary/tossingbot TossingBot is a 2019 system from Princeton and Google in which a robot arm learned to grasp arbitrary objects from clutter and throw them accurately into target bins, achieving pick rates beyond what placing allows. Its key idea, residual physics, combined an analytical ballistic model with a learned network that predicts per-object corrections to release velocity, trained from trial and error. It became a landmark for dynamic manipulation and hybrid model-plus-learning control. Why it matters for physical AI: Exploiting dynamics rather than avoiding them can multiply robot throughput, and TossingBot's residual-physics recipe remains a template for combining known physical models with learned corrections in deployed systems. ### Total Variation Category: Math & Kinematics URL: https://almond.bot/glossary/total-variation Total variation is a measure of the cumulative magnitude of change in a function or signal, defined for a differentiable signal as the integral of the absolute value of its derivative. Total variation regularization penalizes this quantity to suppress noise while preserving sharp edges, a property exploited in image denoising and depth map refinement, and the related total variation distance between probability distributions appears in the analysis of policy update constraints in reinforcement learning. Why it matters for physical AI: Edge-preserving smoothing improves the depth and segmentation maps robots consume, and bounding distributional change between successive policies, a total-variation-style constraint, is the principle behind stable policy-gradient methods used in robot learning. ### Training Data Mixture Category: Foundation Models URL: https://almond.bot/glossary/training-data-mixture A training data mixture is the weighted composition of data sources used to train a model, which for robot foundation models may span teleoperated demonstrations across multiple embodiments, simulation rollouts, human video, and web-scale vision-language corpora. Mixture ratios materially affect capability: cross-embodiment efforts like Open X-Embodiment and RT-X showed careful blending across robot datasets improves transfer, while VLA models co-train on web data to preserve semantic generalization. Why it matters for physical AI: What a robot foundation model can do is largely determined by what it was fed, and mixture design, balancing embodiments, tasks, and modalities, has become as consequential as architecture choices for final policy performance. ### Trajectory Category: Math & Kinematics URL: https://almond.bot/glossary/trajectory A trajectory is a time-parameterized path through a robot's state or configuration space, specifying not just the geometric route but when each state is reached, and hence velocities and accelerations along the way. Trajectories are the common currency between planners and controllers: planners output them, tracking controllers follow them, and their smoothness and dynamic feasibility determine whether execution is possible. In learning contexts, a trajectory also denotes one recorded episode of states and actions. Why it matters for physical AI: Robot learning datasets are collections of trajectories, and the distinction between a path and a timed trajectory matters in practice because learned policies must produce motion that respects real actuator velocity and acceleration limits. ### Trajectory Dataset Category: Data & Benchmarks URL: https://almond.bot/glossary/trajectory-dataset A trajectory dataset is a collection of recorded robot episodes, each pairing time-aligned observations, such as camera images, proprioception, and language annotations, with the actions taken, used to train and evaluate policies. Major examples include Open X-Embodiment, which aggregates over a million trajectories from dozens of robot types, DROID for diverse in-the-wild manipulation, and BridgeData. Standardized formats such as RLDS and LeRobot ease interchange across labs. Why it matters for physical AI: Scale and diversity of trajectory data are the primary drivers of robot foundation model capability, and the field's scarcity of such data relative to text and images is the central bottleneck of physical AI. ### Trajectory Optimization Category: Control URL: https://almond.bot/glossary/trajectory-optimization Trajectory optimization is the computation of a robot motion by minimizing a cost functional, such as time, energy, or control effort, subject to the system's dynamics, actuation limits, and constraints like obstacle avoidance and contact. Methods divide into indirect approaches based on optimality conditions and direct methods, including direct collocation and shooting, that transcribe the problem for nonlinear programming solvers, with DDP and iLQR as prominent dynamic-programming variants. Contact-implicit formulations optimize through contact sequence changes. Why it matters for physical AI: Optimized trajectories provide dynamically feasible, constraint-respecting motion for legged robots and manipulators, serving both as real-time controllers via MPC and as sources of expert supervision for training learned policies. ### Trajectory Tracking Category: Control URL: https://almond.bot/glossary/trajectory-tracking Trajectory tracking is the control problem of making a robot follow a time-parameterized reference trajectory, driving the error between actual and reference states toward zero despite disturbances and model error. Standard solutions include PID and computed-torque control for manipulators, and feedback linearization or nonlinear MPC for mobile robots and quadrotors, with tracking error metrics used to quantify performance. It contrasts with path following, where timing along the route is free. Why it matters for physical AI: Learned policies typically emit setpoints or short reference segments that low-level tracking controllers execute, so tracking bandwidth and accuracy bound the precision of the entire autonomy stack on real hardware. ### Transfer Learning Category: Robot Learning URL: https://almond.bot/glossary/transfer-learning Transfer learning is the reuse of knowledge acquired on a source task, domain, or embodiment to improve learning on a related target, typically by fine-tuning pretrained representations rather than training from scratch. In robotics it spans transferring visual features from web-scale pretraining into control, adapting policies across tasks, sim-to-real transfer, and cross-embodiment transfer between different robots, as demonstrated at scale by the RT-X models on Open X-Embodiment. Why it matters for physical AI: Robots cannot afford to learn every task from zero given the cost of physical data, so transfer from prior models, simulation, and other embodiments is the economic foundation of the entire robot foundation model paradigm. ### Transformer Architecture Category: Foundation Models URL: https://almond.bot/glossary/transformer-architecture The transformer architecture is a neural network design built on self-attention, introduced by Vaswani et al. in 2017, in which every token in a sequence attends to every other, enabling parallel training and scalable modeling of long-range dependencies. Transformers displaced recurrent networks across language and vision and now dominate robotics: policies, world models, and vision-language-action systems tokenize observations and actions into sequences the same attention machinery processes. Why it matters for physical AI: The transformer's uniform treatment of heterogeneous tokens is what lets a single model ingest images, language, proprioception, and actions together, making it the architectural substrate of essentially all current robot foundation models. ### Transformer Policy Category: Robot Learning URL: https://almond.bot/glossary/transformer-policy A transformer policy is a robot control policy implemented as a transformer that maps tokenized observation histories, and often language instructions, to actions. Landmark examples include RT-1, which tokenized camera images and discretized actions for real kitchen tasks, ACT, which decodes action chunks for fine bimanual manipulation, and the autoregressive backbone of VLA models like RT-2 and OpenVLA. Attention over history provides context that reactive feedforward policies lack. Why it matters for physical AI: Transformer policies scale gracefully with data and model size and integrate naturally with pretrained vision-language backbones, which is why they have become the default architecture for imitation-learned manipulation at every scale. ### Transformer Visual Backbone Category: Foundation Models URL: https://almond.bot/glossary/transformer-visual-backbone A transformer visual backbone is a vision encoder based on the Vision Transformer (ViT), which splits an image into patches, embeds them as tokens, and processes them with self-attention to produce visual representations. Pretrained ViT variants, including CLIP encoders, DINOv2, and SigLIP, serve as the perception front-ends of most vision-language and vision-language-action models, with robotics-specific pretraining efforts like R3M and VC-1 targeting manipulation-relevant features. Why it matters for physical AI: The visual backbone determines what a robot policy can perceive before any control learning begins, and reusing internet-pretrained encoders is the main mechanism by which semantic knowledge from the web reaches physical robots. ### Transmission (Mechanical) Category: Hardware URL: https://almond.bot/glossary/transmission-mechanical A transmission is the mechanical stage between a motor and a robot joint that converts speed to torque and relocates or redirects motion, encompassing gearboxes such as planetary, harmonic, and cycloidal drives, as well as belts, ball screws, linkages, and tendons. Transmission choice trades torque density against efficiency, backlash, friction, reflected inertia, and backdrivability: harmonic drives offer zero backlash at high ratios, while low-ratio planetary stages preserve transparency for contact-rich work. Why it matters for physical AI: Transmission properties like friction and reflected inertia are what learned controllers actually feel through the actuators, shaping both the sim-to-real gap and whether a robot can safely absorb impacts and sense contact through its motors. ### Transparency (Teleoperation) Category: Teleoperation URL: https://almond.bot/glossary/transparency-teleoperation Transparency in teleoperation is the fidelity with which a telerobotic system conveys the remote environment to the operator, ideally making the intervening machinery imperceptible so the operator feels as if manipulating the environment directly. Formally it requires the impedance perceived at the leader device to match the remote environment's impedance, a goal that trades off against stability in bilateral force-reflecting systems, especially under communication delay. Why it matters for physical AI: Higher transparency yields more dexterous teleoperation and therefore cleaner, more skillful demonstration data, directly improving the quality of datasets collected for imitation learning at scale. ### Transporter Networks Category: Robot Learning URL: https://almond.bot/glossary/transporter-networks Transporter Networks are a 2020 architecture from Google for vision-based manipulation that rearranges deep visual features to infer spatial displacements: one network attends to a picked region and correlates its features against the scene to score placements, exploiting translational and rotational equivariance. Trained on remarkably few demonstrations, they solved diverse tabletop rearrangement tasks and were released with the Ravens benchmark, influencing subsequent language-conditioned variants like CLIPort. Why it matters for physical AI: Building spatial structure into policy architectures delivers striking sample efficiency for pick-and-place, a lesson that continues to inform the design of data-efficient manipulation models where demonstrations are costly. ### Trapezoidal Velocity Profile Category: Control URL: https://almond.bot/glossary/trapezoidal-velocity-profile A trapezoidal velocity profile is a point-to-point motion trajectory in which velocity ramps up at constant acceleration, holds at a cruise value, and ramps down symmetrically, tracing a trapezoid in the velocity-time plane. It is the default move profile in industrial motion controllers because it is time-optimal under acceleration and velocity limits, though its discontinuous acceleration excites vibration, motivating S-curve profiles with bounded jerk for smoother motion. Why it matters for physical AI: Simple parametric profiles still execute the vast majority of commanded robot moves in industry, and understanding their limits clarifies when learned or optimized trajectories offer genuine value over classical motion generation. ### Tree Search Category: Robot Learning URL: https://almond.bot/glossary/tree-search Tree search is a family of planning methods that explore future possibilities by expanding a tree of states reachable through action sequences, evaluating branches to select the best next action. Variants include classical A* and minimax as well as Monte Carlo Tree Search (MCTS), which balances exploration and exploitation through sampled rollouts and powered AlphaGo and AlphaZero when paired with learned value and policy networks. In robotics, tree search appears in task planning, TAMP, and search over learned world models. Why it matters for physical AI: Deliberate look-ahead over a learned world model offers a route past the reactive limits of behavior-cloned policies, and search-augmented decision-making is an active direction for improving long-horizon reliability in robot systems. ### Trossen Robotics Category: Hardware URL: https://almond.bot/glossary/trossen-robotics Trossen Robotics is a US robotics company known for low-cost research manipulator arms, including the WidowX and ViperX series built on Dynamixel servos with ROS support. Its arms became fixtures of robot learning research: WidowX robots underpin the BridgeData datasets and many Open X-Embodiment contributions, and Trossen commercialized the ALOHA bimanual teleoperation platform as kits used widely for imitation learning data collection. Why it matters for physical AI: Affordable, standardized research arms lowered the cost of entry for manipulation learning and enabled the multi-lab data collection behind several major open datasets, accelerating the broader robot foundation model effort. ### TRPO (Trust Region Policy Optimization) Category: Robot Learning URL: https://almond.bot/glossary/trpo TRPO (Trust Region Policy Optimization) is an on-policy reinforcement learning algorithm introduced by Schulman et al. in 2015 that stabilizes policy-gradient training by constraining each update so the new policy stays within a KL-divergence trust region of the old one, guaranteeing approximate monotonic improvement. It solves a constrained optimization using conjugate gradients and was a milestone for continuous-control learning, though its second-order machinery was largely superseded by the simpler first-order PPO. Why it matters for physical AI: Trust-region thinking, keeping policy updates conservative to avoid catastrophic performance collapse, underlies the PPO-family algorithms that train today's locomotion policies and increasingly fine-tune robot foundation models with reinforcement learning. ### TurtleBot Category: Hardware URL: https://almond.bot/glossary/turtlebot TurtleBot is a family of low-cost, open-source mobile robot platforms created in 2010 at Willow Garage as the reference hardware for learning ROS, now in its fourth generation. Successive versions built on iRobot Create and Kobuki bases, with the widely used TurtleBot3 from ROBOTIS offering a compact differential-drive robot with 2D LiDAR and full ROS 2 support for SLAM, navigation, and multi-robot experiments. Why it matters for physical AI: An accessible standard platform trained a generation of roboticists on navigation and SLAM fundamentals, and shared reference hardware continues to make research reproducible in a field where results are notoriously embodiment-dependent. ### Twist (Spatial Velocity) Category: Math & Kinematics URL: https://almond.bot/glossary/twist-spatial-velocity A twist is the six-dimensional velocity of a rigid body, combining a three-dimensional angular velocity with a three-dimensional linear velocity into a single element of se(3), the Lie algebra of the group of rigid-body motions. Twists unify the description of screw motion: every rigid displacement is the exponential of a twist, a fact exploited by the product-of-exponentials kinematics formulation. The robot Jacobian maps joint velocities to the end-effector twist, and ROS messages carry commanded twists to mobile bases. Why it matters for physical AI: Learned policies commonly output end-effector or base twists as their action representation, and fluency with twists and their frame transformations is essential for correctly wiring policies to real controllers across embodiments. ## Glossary: U ### U-Net Category: Perception URL: https://almond.bot/glossary/u-net U-Net is an encoder-decoder convolutional network architecture introduced by Ronneberger et al. in 2015 for biomedical image segmentation, distinguished by skip connections that pass high-resolution encoder features directly to corresponding decoder stages. The design produces dense per-pixel outputs from limited training data and has become a general-purpose backbone for segmentation, depth estimation, and grasp-affordance prediction, as well as the standard denoising network in diffusion models. Why it matters for physical AI: Dense pixel-wise prediction underlies robot perception tasks from graspability maps to traversability estimation, and U-Net-style denoisers power the diffusion policies now prominent in manipulation learning. ### UART Category: Hardware URL: https://almond.bot/glossary/uart UART (Universal Asynchronous Receiver-Transmitter) is a serial communication peripheral that exchanges data over two wires without a shared clock, framing each byte with start and stop bits at a preagreed baud rate. It is among the simplest and most ubiquitous interfaces in embedded robotics, linking microcontrollers to GPS receivers, IMUs, servo controllers, and debug consoles, though its point-to-point nature and modest speeds cede demanding applications to buses like CAN and Ethernet. Why it matters for physical AI: Low-level sensor and actuator plumbing still runs through interfaces like UART, and understanding their latency, bandwidth, and framing limits matters when integrating hardware whose data feeds real-time control loops. ### UMI (Universal Manipulation Interface) Category: Teleoperation URL: https://almond.bot/glossary/umi UMI (Universal Manipulation Interface) is a handheld, robot-free data collection system introduced by Chi et al. in 2024, consisting of a portable gripper with a wrist-mounted GoPro fisheye camera that records manipulation demonstrations anywhere a person can carry it. Its SLAM-based pose recovery, wide-view observations, and latency-matched policy interface let demonstrations collected in the wild transfer to real robot arms, with the follow-up Mobile UMI extending the recipe to whole-body mobile manipulation. Why it matters for physical AI: Decoupling demonstration collection from robot hardware removes a principal bottleneck in scaling manipulation datasets, letting data grow at the pace of people with handheld devices rather than instrumented robot cells. ### Uncertainty Quantification Category: Robot Learning URL: https://almond.bot/glossary/uncertainty-quantification Uncertainty quantification is the estimation of how confident a model's predictions are, conventionally divided into aleatoric uncertainty, inherent in noisy data, and epistemic uncertainty, stemming from limited knowledge that more data could reduce. Practical methods include deep ensembles, Monte Carlo dropout, Bayesian neural networks, and conformal prediction, which wraps any model with distribution-free coverage guarantees. Well-calibrated uncertainty lets a system know when it does not know. Why it matters for physical AI: Robots acting on wrong-but-confident predictions cause physical damage, so calibrated uncertainty is the trigger for safe fallbacks, human handoffs, and targeted data collection in deployed learning systems. ### Underactuated Gripper Category: Hardware URL: https://almond.bot/glossary/underactuated-gripper An underactuated gripper is a robotic hand with fewer actuators than mechanical degrees of freedom, whose fingers close through linkages, tendons, or compliant joints that let them passively conform to an object's shape. Designs such as the SDM Hand and Robotiq adaptive grippers achieve robust power grasps of varied objects with a single motor, trading precise fingertip control for mechanical intelligence that tolerates pose uncertainty without sensing. Why it matters for physical AI: Passive conformity absorbs the perception and control errors that learned grasping systems inevitably make, which is why underactuated designs remain a pragmatic choice for reliable picking in warehouses and unstructured settings. ### Underactuation Category: Control URL: https://almond.bot/glossary/underactuation Underactuation is the condition in which a mechanical system has fewer independent actuators than degrees of freedom, so some directions of motion cannot be commanded directly and the dynamics must be exploited rather than overridden. Walking robots in flight or single-support phases, quadrotors, the Acrobot and cart-pole, and objects manipulated through contact are all underactuated. Control relies on trajectory optimization, partial feedback linearization, and limit-cycle stability rather than exact tracking. Why it matters for physical AI: Legged locomotion and dynamic manipulation are irreducibly underactuated, so both classical controllers and learned policies for these systems must reason about momentum and contact timing instead of simply servoing every degree of freedom. ### Underwater Robot Category: Applications URL: https://almond.bot/glossary/underwater-robot An underwater robot is a marine vehicle that operates submerged, spanning tethered remotely operated vehicles (ROVs) piloted from the surface and untethered autonomous underwater vehicles (AUVs) that execute missions independently. Applications include offshore infrastructure and hull inspection, oceanographic survey, pipeline monitoring, and deep-sea exploration. The medium imposes distinctive constraints: radio waves attenuate rapidly, forcing acoustic communication and navigation, while pressure, corrosion, and hydrodynamic disturbances shape every design decision. Why it matters for physical AI: GPS-denied, communication-limited environments demand high onboard autonomy, making underwater robotics a proving ground for self-contained perception and decision-making that cannot lean on remote compute or human oversight. ### Uneven Terrain Locomotion Category: Locomotion URL: https://almond.bot/glossary/uneven-terrain-locomotion Uneven terrain locomotion is the problem of legged or wheeled robots traversing ground that is irregular, sloped, discontinuous, or deformable, such as stairs, rubble, trails, and vegetation. Modern legged approaches train reinforcement learning policies over procedurally generated terrains in massively parallel simulation, often with teacher-student distillation from privileged terrain information, as in work on ANYmal, and perceptive variants fuse elevation maps or depth vision with proprioception to place feet deliberately. Why it matters for physical AI: Terrain robustness is what separates robots confined to flat warehouse floors from those useful in construction, inspection, agriculture, and disaster response, and it remains the benchmark capability for legged platforms. ### Unimate Category: Hardware URL: https://almond.bot/glossary/unimate Unimate was the first industrial robot, a hydraulically powered programmable arm invented by George Devol, commercialized with Joseph Engelberger through Unimation, and installed on a General Motors die-casting line in New Jersey in 1961. It replayed step-by-step programs stored on a magnetic drum to unload hot castings, work hazardous to humans, and its success launched the industrial robotics industry and the automation of automotive manufacturing worldwide. Why it matters for physical AI: The arc from Unimate's fixed replay of recorded waypoints to today's learned, vision-guided policies frames how far robot autonomy has come, and its deployment logic, taking over dangerous repetitive work, still drives adoption. ### Unitree G1 Category: Hardware URL: https://almond.bot/glossary/unitree-g1 The Unitree G1 is a compact humanoid robot introduced by Unitree Robotics in 2024, standing roughly 1.3 meters tall and weighing about 35 kilograms, with a starting price near 16,000 US dollars that undercut prior humanoids by an order of magnitude. It offers over 23 joint degrees of freedom, optional three-fingered hands, and 3D LiDAR plus depth sensing, and it has become a widespread research platform for whole-body reinforcement learning and teleoperated data collection. Why it matters for physical AI: Affordable humanoid hardware moved whole-body learning research from a handful of well-funded labs to the broader community, accelerating progress on humanoid locomotion, teleoperation pipelines, and foundation-model control. ### Unitree Go2 Category: Hardware URL: https://almond.bot/glossary/unitree-go2 The Unitree Go2 is a consumer and research quadruped robot released by Unitree Robotics in 2023, weighing about 15 kilograms with 12 actuated joints, an onboard 4D LiDAR, and a starting price under 2,000 US dollars for the base model. Research-oriented EDU variants expose low-level joint control and support secondary compute, making the Go2 one of the most widely used platforms for legged reinforcement learning coursework and sim-to-real experiments. Why it matters for physical AI: Cheap, robust quadrupeds with open low-level interfaces made legged locomotion research reproducible at scale, and fleets of such platforms now serve as testbeds for perceptive locomotion policies trained in massively parallel simulation. ### Unitree H1 Category: Hardware URL: https://almond.bot/glossary/unitree-h1 The Unitree H1 is a full-size humanoid robot released by Unitree Robotics in 2023, standing about 1.8 meters tall and weighing roughly 47 kilograms, with high-torque electric joint actuators and 3D LiDAR sensing. It set an official Guinness World Record for humanoid running speed at 3.3 meters per second and demonstrated reinforcement-learning-trained feats including a standing backflip without hydraulics, establishing electric humanoids as dynamically capable research platforms. Why it matters for physical AI: Demonstrations of dynamic, learned whole-body control on commercially available hardware validated the simulation-trained reinforcement learning recipe for humanoids and intensified the race toward deployable general-purpose bipeds. ### Universal Robots UR5 Category: Hardware URL: https://almond.bot/glossary/universal-robots-ur5 The Universal Robots UR5 is a 6-DoF collaborative robot arm introduced in 2008 by Danish manufacturer Universal Robots, with a 5 kilogram payload and 850 millimeter reach, and among the first arms designed to work alongside people without safety cages by limiting force and power. Its successor UR5e continues the line. Simple programming, safety certification, and moderate cost made the UR family the best-selling collaborative arms and a ubiquitous fixture in manipulation research labs. Why it matters for physical AI: The UR5's ubiquity made it a de facto standard embodiment in manipulation datasets and benchmarks, and its collaborative safety model prefigured the human-adjacent deployment context that learned manipulation policies now target. ### Unstructured Environment Category: Applications URL: https://almond.bot/glossary/unstructured-environment An unstructured environment is a setting whose layout, objects, and conditions are not engineered or fixed for the robot, in contrast to structured industrial cells with fixtures, known part poses, and controlled lighting. Homes, hospitals, construction sites, farms, and disaster zones are canonical examples, presenting clutter, occlusion, deformable and novel objects, variable illumination, and people. Operating there requires perception-driven, adaptive behavior rather than replayed motions. Why it matters for physical AI: The economic promise of robot foundation models rests precisely on unstructured settings that defeat traditional automation, making robustness to open-world variability the defining test of general-purpose robot intelligence. ### Unsupervised Learning Category: Robot Learning URL: https://almond.bot/glossary/unsupervised-learning Unsupervised learning is the extraction of structure from data without human-provided labels, encompassing clustering, dimensionality reduction, density estimation, and generative modeling. In modern practice it is often realized as self-supervised learning, where supervision is manufactured from the data itself, as in contrastive methods, masked reconstruction, and next-token prediction. Robotics applications include learning visual representations from unlabeled video, skill discovery, and world models trained on raw interaction streams. Why it matters for physical AI: Physical interaction data is expensive to label at scale, so extracting representations, skills, and dynamics from unlabeled robot and human video is essential to feeding data-hungry foundation models economically. ### Upper-Body Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/upper-body-teleoperation Upper-body teleoperation is the remote control of a humanoid or bimanual robot's arms, hands, torso, and head by mapping a human operator's corresponding upper-body motion, captured through VR headsets and hand tracking, motion-capture suits, or exoskeletons, onto the robot. Retargeting must reconcile differing kinematics and joint limits, and systems such as Open-TeleVision and AnyTeleop stream head-tracked stereo vision back to the operator for immersive, precise bimanual manipulation. Why it matters for physical AI: Humanoid manipulation data is overwhelmingly collected through upper-body teleoperation, so retargeting quality and operator immersion directly bound the dexterity that imitation-learned humanoid policies can acquire. ### URDF (Unified Robot Description Format) Category: Software & Middleware URL: https://almond.bot/glossary/urdf URDF (Unified Robot Description Format) is an XML format from the ROS ecosystem for describing a robot's kinematic tree, specifying links with their inertial, visual, and collision properties and the joints connecting them. It is consumed by visualization, kinematics, motion planning, and simulation tools, commonly generated via xacro macros. Its tree-only topology cannot express closed kinematic chains, a limitation addressed by successors and alternatives such as SDF and MJCF. Why it matters for physical AI: A robot's URDF is its identity card across simulators, planners, and learning frameworks, and accurate descriptions, especially inertial parameters, are quietly critical to sim-to-real transfer of learned controllers. ### USD (Universal Scene Description) Category: Simulation URL: https://almond.bot/glossary/usd USD (Universal Scene Description) is an open-source framework originally developed by Pixar for composing, editing, and interchanging hierarchical 3D scenes, built around layered, non-destructive composition of scene description. In robotics it is the native scene format of NVIDIA Isaac Sim and the Omniverse platform, where OpenUSD schemas describe robots, environments, physics properties, and sensors for photorealistic simulation and synthetic data generation. Why it matters for physical AI: Simulation-generated training data depends on rich, composable scene descriptions, and USD's layering makes it practical to procedurally vary environments at the scale that domain randomization and synthetic data pipelines demand. ### UV Disinfection Robot Category: Applications URL: https://almond.bot/glossary/uv-disinfection-robot A UV disinfection robot is a mobile robot that navigates indoor spaces emitting ultraviolet-C light, typically at 254 nanometers, which inactivates bacteria and viruses by damaging their nucleic acids. Deployed in hospitals, airports, and offices, and adopted widely during the COVID-19 pandemic with systems such as those from UVD Robots, they combine autonomous navigation with strict safety interlocks, since germicidal UV-C harms human skin and eyes, requiring vacancy detection before operation. Why it matters for physical AI: Disinfection robots illustrate autonomy delivering value in service environments today, and their requirement to guarantee no human exposure exemplifies the safety-assurance problems any robot operating in shared spaces must solve. ### UWB Positioning Category: Navigation & SLAM URL: https://almond.bot/glossary/uwb-positioning UWB (ultra-wideband) positioning is a radio localization technique that measures distances between tags and fixed anchors from the time of flight of nanosecond-scale pulses spread over greater than 500 megahertz of bandwidth, yielding indoor accuracy of roughly 10 to 30 centimeters. Its wide bandwidth resolves multipath far better than WiFi or Bluetooth ranging. Robots use UWB for GPS-denied localization, multi-robot relative positioning, and person following, often fused with odometry and inertial sensing. Why it matters for physical AI: Reliable infrastructure-based positioning gives warehouse and service robots a drift-free localization anchor where GPS fails and visual methods struggle, simplifying fleet coordination in large, repetitive indoor spaces. ## Glossary: V ### Vacuum Gripper Category: Hardware URL: https://almond.bot/glossary/vacuum-gripper A vacuum gripper is an end effector that holds objects with negative air pressure applied through one or more suction cups, generated by a vacuum pump or a compressed-air Venturi ejector. It excels on flat, smooth, non-porous surfaces such as boxes, glass, and sheet metal, and is the dominant gripper type in logistics picking and palletizing. Grip reliability depends on seal quality, cup geometry and material, and object porosity. Why it matters for physical AI: Suction simplifies the grasping problem to predicting a single contact point, which is why many learned bin-picking systems output suction poses and why warehouse deployments of robot picking overwhelmingly rely on vacuum end effectors. ### Vacuum Pump Category: Hardware URL: https://almond.bot/glossary/vacuum-pump A vacuum pump is a device that removes air from a sealed volume to create the negative pressure that drives suction-based end effectors. Robotic systems use electric diaphragm or rotary vane pumps for continuous vacuum, or Venturi ejectors that generate vacuum from compressed air with no moving parts. Key specifications include achievable vacuum level, flow rate (which determines recovery after leaks), and response time for gripping and release cycles. Why it matters for physical AI: Pump sizing determines pick reliability on porous or irregular objects, and vacuum sensing (detecting whether a seal was achieved) provides a cheap, robust grasp-success signal for closed-loop picking and for automatically labeling manipulation data. ### Value Function Category: Robot Learning URL: https://almond.bot/glossary/value-function A value function is an estimate of the expected cumulative reward an agent will accumulate from a given state, written V(s), or from a state-action pair, written Q(s, a), while following a particular policy. Value functions satisfy the Bellman equation and are the central object of dynamic programming, temporal-difference learning, and actor-critic reinforcement learning, where a learned critic evaluates actions proposed by the policy. Why it matters for physical AI: Value estimates let robots trained with reinforcement learning trade off long-horizon consequences of actions, and pretrained value functions are increasingly used to score, filter, or steer imitation-learned policies at deployment time. ### Value Iteration Category: Robot Learning URL: https://almond.bot/glossary/value-iteration Value iteration is a dynamic programming algorithm that computes the optimal value function of a Markov decision process by repeatedly applying the Bellman optimality backup to every state until the values converge, after which the optimal policy is extracted greedily. It requires a known transition model and an enumerable state space, so in robotics it is applied to discretized problems such as grid-based navigation and serves as the conceptual foundation for approximate and learned variants. Why it matters for physical AI: Grid-world value iteration still powers practical costmap-based planners, and its fixed-point structure underlies deep RL methods, making it essential background for understanding how learned critics and planners behave. ### Value-Implicit Pre-training (VIP) Category: Robot Learning URL: https://almond.bot/glossary/value-implicit-pretraining Value-Implicit Pre-training (VIP) is a self-supervised representation learning method, introduced by Ma et al. in 2022, that trains a visual encoder on large-scale human video from Ego4D so that distances in its embedding space behave like a goal-conditioned value function. The resulting representation provides dense, zero-shot reward signals for downstream robot manipulation tasks without requiring any robot-specific or task-specific reward engineering. Why it matters for physical AI: Reward specification is a major bottleneck for reinforcement learning on real robots; representations like VIP show how passive human video can supply transferable value estimates and dense rewards for physical tasks. ### Variable Stiffness Actuator (VSA) Category: Hardware URL: https://almond.bot/glossary/variable-stiffness-actuator A variable stiffness actuator (VSA) is an actuator whose mechanical compliance can be adjusted independently of its output position, typically by using a second motor to reconfigure internal elastic elements. Unlike a series elastic actuator with a fixed spring, a VSA can stiffen for precise positioning and soften for impact absorption or safe human contact within the same task, storing and releasing elastic energy for explosive motions. Why it matters for physical AI: Physical compliance that is tunable in hardware absorbs impacts faster than any control loop can react, which matters for robots working in contact-rich, human-populated environments where collisions are inevitable. ### Variational Autoencoder (VAE) Category: Robot Learning URL: https://almond.bot/glossary/variational-autoencoder A variational autoencoder (VAE) is a generative model, introduced by Kingma and Welling in 2013, that learns a probabilistic latent representation of data by jointly training an encoder and decoder to maximize the evidence lower bound (ELBO). Conditional VAEs (CVAEs) are widely used in robot learning to model multimodal action distributions; the ACT policy from the ALOHA project, for example, uses a CVAE over action chunks. Why it matters for physical AI: Demonstration data is inherently multimodal because humans solve the same task in different ways, and VAE-style latent variables let imitation policies capture that variability instead of averaging conflicting demonstrations into invalid actions. ### Variational Inference Category: Robot Learning URL: https://almond.bot/glossary/variational-inference Variational inference is a technique for approximating an intractable posterior distribution by optimizing the parameters of a simpler distribution to minimize its KL divergence from the true posterior, usually by maximizing the evidence lower bound (ELBO). It converts Bayesian inference into an optimization problem that scales to deep networks, and it underlies variational autoencoders, latent-variable policies, and several model-based reinforcement learning methods. Why it matters for physical AI: Robots must reason under uncertainty from noisy, partial observations, and variational methods provide the tractable machinery behind learned state estimators, world models such as Dreamer, and latent-variable policy architectures. ### Vector Quantization (VQ) Category: Robot Learning URL: https://almond.bot/glossary/vector-quantization Vector quantization (VQ) is the mapping of continuous vectors to the nearest entry in a learned discrete codebook, popularized in deep learning by the VQ-VAE. In robot learning it is used to tokenize continuous observations and actions into discrete symbols so that transformer policies can model behavior with categorical prediction heads; VQ-BeT, for example, quantizes action chunks to capture multimodal demonstration data. Why it matters for physical AI: Discretizing actions lets robot policies borrow the full machinery of language models, including next-token prediction and autoregressive sampling, which has become a dominant recipe in vision-language-action architectures. ### Velocity Control Category: Control URL: https://almond.bot/glossary/velocity-control Velocity control is a control mode in which the commanded quantity is joint or end-effector velocity rather than position or torque, executed by an inner servo loop that tracks the velocity setpoint. It is the native interface for wheeled mobile bases and is central to resolved-rate motion control, where a desired Cartesian twist is mapped to joint velocities through the Jacobian pseudoinverse. Why it matters for physical AI: Streaming velocity commands is the standard interface for teleoperation and visual servoing, and many learned policies output velocity actions because they are smooth, bounded, and safer to execute than raw torques. ### Velocity Kinematics Category: Math & Kinematics URL: https://almond.bot/glossary/velocity-kinematics Velocity kinematics is the study of the relationship between a robot's joint velocities and the linear and angular velocities of its links, expressed through the manipulator Jacobian, which maps joint rates to the end-effector twist. The Jacobian's transpose maps end-effector wrenches back to joint torques, and configurations where the Jacobian loses rank are kinematic singularities at which certain motions become infeasible. Why it matters for physical AI: Every Cartesian controller, teleoperation retargeter, and singularity-avoidance scheme runs on velocity kinematics; policies that command end-effector twists depend on it to translate their outputs into joint motion. ### Velocity Obstacles Category: Navigation & SLAM URL: https://almond.bot/glossary/velocity-obstacles A velocity obstacle is the set of velocities that would cause a moving agent to collide with a moving obstacle at some future time, assuming the obstacle maintains its current velocity; selecting velocities outside this set guarantees collision-free motion. Introduced by Fiorini and Shiller in 1998, the concept was extended to reciprocal velocity obstacles (RVO) and ORCA for decentralized collision avoidance among many agents. Why it matters for physical AI: Mobile robots and fleets operating around people need reactive, provable collision avoidance at high control rates, and velocity-obstacle methods remain the workhorse for multi-robot and crowd navigation. ### Vibration Damping Category: Control URL: https://almond.bot/glossary/vibration-damping Vibration damping is the attenuation of unwanted oscillation in a mechanical system by dissipating or canceling vibratory energy, achieved passively with damping materials and isolators or actively with sensors and actuators in closed loop. In robotics it is critical for lightweight arms, long-reach and cable-driven mechanisms, and high-speed gantries, where structural flexibility causes residual vibration; input shaping and active damping control are common countermeasures. Why it matters for physical AI: Lightweight, low-cost arms flex noticeably, and uncontrolled vibration degrades precision, blurs camera observations, and injects noise into demonstration data, so damping directly affects both task performance and learned-policy quality. ### Video Prediction Model Category: Foundation Models URL: https://almond.bot/glossary/video-prediction-model A video prediction model is a generative model that forecasts future video frames conditioned on past frames and, optionally, on actions or language. Action-conditioned video prediction enabled early visual model-predictive control in the Visual Foresight line of work, and modern large-scale video generators are increasingly treated as learned world simulators from which robot policies can plan or be distilled. Why it matters for physical AI: A model that predicts how scenes evolve under actions is effectively a physics simulator learned from data, offering a path to planning and policy training that scales with web video rather than hand-built simulation. ### Vineyard Robot Category: Applications URL: https://almond.bot/glossary/vineyard-robot A vineyard robot is an agricultural mobile robot designed to operate between vine rows, performing tasks such as precision spraying, mowing, canopy and yield monitoring, and harvest support. The structured row geometry of vineyards suits autonomous navigation, but dense canopy limits GNSS reliability, pushing these platforms toward LiDAR- and vision-based row following, and steep or uneven terrain drives tracked and articulated chassis designs. Why it matters for physical AI: Agriculture's labor shortages and chemical-reduction targets make vineyards a proving ground for field robotics, exercising perception, traversability estimation, and manipulation under seasonal appearance change that indoor benchmarks never capture. ### ViperX 300 Category: Hardware URL: https://almond.bot/glossary/viperx-300 The ViperX 300 is a low-cost 6-DOF research manipulator from Trossen Robotics, built from Dynamixel smart servos with roughly 750 mm reach and 750 g payload. It became widely known as the follower arm in the ALOHA bimanual teleoperation platform, where a pair of ViperX 300 arms is puppeteered via smaller WidowX 250 leader arms to collect fine manipulation demonstrations. Why it matters for physical AI: Affordable, hackable arms like the ViperX lowered the cost of bimanual imitation-learning research by an order of magnitude, seeding the wave of open hardware platforms on which much current manipulation data is collected. ### Virtual Fixture Category: Teleoperation URL: https://almond.bot/glossary/virtual-fixture A virtual fixture is a software-defined constraint overlaid on a teleoperated or shared-control robot that guides the operator's motion, either blocking entry into a region (forbidden-region fixture) or attracting motion along a path or surface (guidance fixture), often rendered haptically. Introduced by Rosenberg in the early 1990s, virtual fixtures are widely used in surgical robotics and precision teleoperation. Why it matters for physical AI: Blending human intent with machine-enforced constraints raises teleoperation precision and safety, and the same shared-autonomy machinery underlies systems where a learned policy and a human operator control a robot together. ### Virtual Model Control Category: Control URL: https://almond.bot/glossary/virtual-model-control Virtual model control is a control technique that emulates imaginary mechanical components, such as springs and dampers, attached between points on a robot and its environment, computing the joint torques those components would exert if they physically existed. Developed by Pratt and colleagues at the MIT Leg Laboratory for bipedal walking, it provides an intuitive, low-dimensional way to specify compliant, force-based behaviors without explicit trajectory planning. Why it matters for physical AI: Specifying behavior as virtual springs and dampers remains a core idiom for legged locomotion and compliant manipulation, and it connects directly to the impedance controllers that keep contact-rich learned behaviors stable on hardware. ### Vision Encoder Category: Foundation Models URL: https://almond.bot/glossary/vision-encoder A vision encoder is the neural network component that maps raw images into compact feature embeddings consumed by downstream policy, planning, or reasoning modules. Common choices in robot learning include convolutional backbones such as ResNet and transformer backbones such as ViT, often pretrained with objectives like CLIP, SigLIP, or DINOv2; vision-language-action models typically fuse one or more vision encoders with a language-model backbone. Why it matters for physical AI: Encoder choice and pretraining largely determine what a robot policy can perceive, and reusing frozen web-pretrained encoders is one of the main mechanisms by which internet-scale visual knowledge transfers into physical skills. ### Vision Transformer (ViT) Category: Foundation Models URL: https://almond.bot/glossary/vision-transformer The Vision Transformer (ViT) is a neural architecture, introduced by Dosovitskiy et al. in 2020, that applies a standard transformer to images by splitting them into fixed-size patches, linearly embedding each patch, and processing the resulting token sequence with self-attention. Given large-scale pretraining, ViTs match or surpass convolutional networks and now serve as the default vision backbone for CLIP, DINOv2, SAM, and most vision-language models. Why it matters for physical AI: Because ViTs share the transformer substrate with language models, they make unified multimodal architectures natural, which is why nearly every modern robot foundation model tokenizes camera input through a ViT-family encoder. ### Vision-Based Tactile Sensing Category: Perception URL: https://almond.bot/glossary/vision-based-tactile-sensing Vision-based tactile sensing is a sensing approach in which a camera inside the fingertip images the deformation of an illuminated elastomer pad pressed against an object, recovering high-resolution contact geometry, force distribution, and slip. GelSight established the technique, and compact open-hardware designs such as Meta's DIGIT made it practical to mount on standard grippers; photometric stereo on the pad yields detailed surface normal maps. Why it matters for physical AI: Contact-rich skills like insertion, cable routing, and in-hand reorientation are hard to learn from vision alone, and camera-based tactile skins deliver dense touch data in a form that existing visual policy architectures can consume directly. ### Vision-Language Model (VLM) Category: Foundation Models URL: https://almond.bot/glossary/vision-language-model A vision-language model (VLM) is a neural network trained on paired image and text data to jointly represent and reason over visual and linguistic content, supporting tasks such as captioning, visual question answering, and open-vocabulary grounding. Contrastive models like CLIP align images and text in a shared embedding space, while generative VLMs such as PaLI, LLaVA, and GPT-4-class multimodal models produce free-form text conditioned on images. Why it matters for physical AI: VLMs give robots open-vocabulary perception and commonsense scene understanding for free, and they are the pretrained substrate from which vision-language-action models are fine-tuned on robot data. ### Vision-Language-Action Model (VLA) Category: Foundation Models URL: https://almond.bot/glossary/vision-language-action-model A vision-language-action model (VLA) is a foundation model that maps camera observations and natural-language instructions directly to robot actions, typically created by fine-tuning a pretrained vision-language model on large robot demonstration datasets. RT-2 established the paradigm by expressing actions as text tokens; subsequent models such as OpenVLA and Pi0 train on cross-embodiment corpora like Open X-Embodiment and add continuous action heads such as flow matching. Why it matters for physical AI: VLAs are the leading candidate architecture for general-purpose robot policies, transferring web-scale semantic knowledge into control and improving predictably with more diverse robot data, which reframes manipulation as a data-scaling problem. ### Visual Foundation Model Category: Foundation Models URL: https://almond.bot/glossary/visual-foundation-model A visual foundation model is a large vision network pretrained on web-scale image data whose representations transfer broadly across downstream tasks without task-specific training. Prominent examples include DINOv2 for self-supervised features, SAM (Segment Anything) for promptable segmentation, and CLIP's image tower for language-aligned embeddings. Robotics pipelines use them frozen or lightly fine-tuned for detection, segmentation, correspondence, and as policy encoders. Why it matters for physical AI: Robot datasets are small compared to internet imagery, so leaning on visual foundation models for perception lets manipulation and navigation systems generalize to novel objects and scenes their training data never covered. ### Visual Imitation Category: Robot Learning URL: https://almond.bot/glossary/visual-imitation Visual imitation is imitation learning in which the policy consumes raw camera images rather than privileged state, and, in a broader sense, learning behaviors from visual observation of another agent, including humans, without access to expert actions. The first form dominates modern manipulation through methods like ACT and Diffusion Policy; the second, learning from human video, must additionally bridge the embodiment gap between demonstrator and robot. Why it matters for physical AI: Cameras are the most scalable sensor for collecting demonstrations, and progress in visual imitation determines how directly the vast supply of human manipulation video can be converted into robot skills. ### Visual Odometry Category: Navigation & SLAM URL: https://almond.bot/glossary/visual-odometry Visual odometry (VO) is the estimation of a camera's egomotion by tracking visual features or intensity patterns across consecutive frames and integrating the resulting incremental pose changes. Monocular VO recovers trajectory up to scale, while stereo and visual-inertial variants (VIO) resolve scale and improve robustness by fusing IMU measurements. Like all dead reckoning, it accumulates drift, which SLAM systems correct through loop closure. Why it matters for physical AI: Drift-bounded egomotion from cheap cameras is foundational for drones, legged robots, and mobile manipulators operating where GPS fails, and VIO is the de facto state estimation layer beneath most autonomous navigation stacks. ### Visual Place Recognition Category: Navigation & SLAM URL: https://almond.bot/glossary/visual-place-recognition Visual place recognition (VPR) is the task of determining whether a current camera view corresponds to a previously visited place, typically by matching global image descriptors, such as NetVLAD embeddings, against a database of mapped locations. Robust VPR must handle viewpoint shift, lighting and seasonal change, and perceptual aliasing, where distinct places look alike. It supplies loop-closure candidates in SLAM and coarse relocalization after tracking failure. Why it matters for physical AI: Long-term autonomy hinges on recognizing previously mapped space despite appearance change; without reliable place recognition, maps drift apart and robots cannot relocalize after being moved or powered down. ### Visual Servoing Category: Control URL: https://almond.bot/glossary/visual-servoing Visual servoing is closed-loop robot control in which the feedback signal comes directly from camera measurements. Image-based visual servoing (IBVS) regulates errors defined in image-feature coordinates using the interaction matrix, while position-based visual servoing (PBVS) regulates errors in the estimated 3D pose of the target; hybrid schemes combine both. Continuous visual feedback confers robustness to calibration error and target motion. Why it matters for physical AI: Closing the loop through vision is what lets robots achieve millimeter-level alignment with imprecise kinematics, and modern learned policies that react to images at high rate are effectively visual servoing generalized beyond hand-designed features. ### Visual SLAM Category: Navigation & SLAM URL: https://almond.bot/glossary/visual-slam Visual SLAM is simultaneous localization and mapping using cameras as the primary sensor, jointly estimating the robot's trajectory and a 3D map of the environment from image streams. Feature-based systems such as ORB-SLAM track sparse keypoints and optimize a pose graph with loop closure, while direct and dense methods operate on pixel intensities; visual-inertial fusion adds robustness to fast motion and low texture. Why it matters for physical AI: Cameras cost dollars while LiDAR costs hundreds, so visual SLAM is the enabling technology for affordable navigation in consumer robots, drones, and AR headsets, and it supplies the localization layer for mobile manipulation. ### Visuomotor Policy Category: Robot Learning URL: https://almond.bot/glossary/visuomotor-policy A visuomotor policy is a control policy that maps raw camera observations, usually combined with proprioception, directly to motor commands, trained end to end rather than composed from separate perception, planning, and control modules. Levine et al.'s 2016 work on end-to-end training of deep visuomotor policies established the approach; contemporary instances include Diffusion Policy, ACT, and the action heads of vision-language-action models. Why it matters for physical AI: End-to-end visuomotor learning removes the brittle hand-engineered interfaces between perception and control, and it is the core policy abstraction underneath essentially every current robot foundation model effort. ### Voice Coil Actuator Category: Hardware URL: https://almond.bot/glossary/voice-coil-actuator A voice coil actuator is a direct-drive electromagnetic actuator in which a coil moves within a permanent-magnet field, producing force proportional to current over a short stroke, with no cogging, backlash, or gearing. Its low moving mass and clean force-current relationship give it exceptional bandwidth and force fidelity, which is why it appears in haptic interfaces, fast fine-positioning stages, high-speed grippers, and optical focus mechanisms. Why it matters for physical AI: High-bandwidth, backdrivable force sources are essential for rendering haptic feedback in teleoperation and for delicate contact tasks, where geared actuators are too stiff and slow to react to force transients. ### Voxel Grid Category: Perception URL: https://almond.bot/glossary/voxel-grid A voxel grid is a discretization of 3D space into regular cubic cells (voxels), each storing quantities such as occupancy probability, truncated signed distance, color, or learned features. Voxel grids underpin occupancy mapping frameworks like OctoMap, TSDF-based reconstruction, collision checking for motion planning, and the voxel-downsampling filters used to regularize point clouds. Memory grows cubically with resolution, motivating sparse structures such as octrees. Why it matters for physical AI: Explicit volumetric maps give manipulation and navigation systems queryable free-space and surface geometry, and voxelized scene features increasingly feed 3D policy architectures that reason beyond single camera views. ### Voxel Hashing Category: Perception URL: https://almond.bot/glossary/voxel-hashing Voxel hashing is a sparse volumetric data structure, introduced by Niessner et al. in 2013, that stores only voxel blocks near observed surfaces in a spatial hash table instead of allocating a dense grid over the whole scene. It enables real-time TSDF fusion and large-scale 3D reconstruction at fine resolution with bounded memory, and it remains a standard component in GPU reconstruction pipelines. Why it matters for physical AI: Reconstructing workspaces at millimeter resolution in real time is only tractable with sparse volumetric storage, making voxel hashing a key enabler for on-robot dense mapping and scene-level manipulation planning. ### VoxPoser Category: Foundation Models URL: https://almond.bot/glossary/voxposer VoxPoser is a zero-shot manipulation method from Huang et al. (2023) in which a large language model writes code that composes vision-language-model queries into 3D voxel value maps encoding task affordances and constraints, and a motion planner then synthesizes end-effector trajectories by optimizing over those maps. It executes free-form language instructions without any task-specific policy training or demonstrations. Why it matters for physical AI: VoxPoser demonstrated that foundation models can ground language into executable spatial plans without robot training data, an influential template for zero-shot manipulation and for hybrids of learned reasoning with classical planning. ### VR Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/vr-teleoperation VR teleoperation is the control of a robot through a virtual-reality headset and tracked hand controllers or hand tracking, with the operator's head and hand motions retargeted to the robot's cameras and end effectors, often with stereoscopic video feedback for depth perception. Consumer headsets such as the Meta Quest have made it a widespread low-cost interface, and systems like Open-TeleVision and AnyTeleop popularized it for bimanual data collection. Why it matters for physical AI: Immersive teleoperation is one of the highest-throughput ways to collect dexterous bimanual demonstrations, directly feeding the demonstration-hungry imitation pipelines behind current robot foundation models. ## Glossary: W ### Walking Category: Locomotion URL: https://almond.bot/glossary/walking Walking is legged locomotion in which at least one foot remains in ground contact at all times, distinguishing it from running, which includes flight phases. Robot walking control spans ZMP-based trajectory generation for humanoids, capture-point and model-predictive controllers built on reduced models like the linear inverted pendulum, and, increasingly, reinforcement-learned policies trained in simulation and transferred to hardware. Why it matters for physical AI: Robust walking is the gateway capability for humanoids and quadrupeds entering human environments, and the sim-to-real recipes proven on locomotion now inform how contact-rich skills of all kinds are trained. ### Warehouse Robot Category: Applications URL: https://almond.bot/glossary/warehouse-robot A warehouse robot is a robot deployed in logistics facilities for tasks such as goods-to-person transport, shelf and tote movement, piece picking, sorting, and palletizing. Amazon's acquisition of Kiva Systems in 2012 catalyzed the category; today fleets of autonomous mobile robots coordinate through centralized traffic control, while vision-guided arms handle picking from bins and conveyors. Why it matters for physical AI: Warehouses are the largest commercial proving ground for embodied AI, generating enormous volumes of real manipulation and navigation data under hard throughput and reliability constraints that push learned systems beyond lab conditions. ### Wavelet Transform Category: Math & Kinematics URL: https://almond.bot/glossary/wavelet-transform The wavelet transform is a signal-processing technique that decomposes a signal into components localized in both time and frequency using scaled and shifted copies of a basis wavelet, unlike the Fourier transform, which localizes only in frequency. In robotics it is applied to vibration and fault analysis, tactile and force signal processing, and multiresolution image and terrain analysis. Why it matters for physical AI: Contact events, slips, and mechanical faults are transient signals buried in noisy sensor streams, and time-frequency decompositions remain effective features for detecting them in real time on resource-limited hardware. ### Waypoint Category: Navigation & SLAM URL: https://almond.bot/glossary/waypoint A waypoint is an intermediate target position or pose along a robot's route or trajectory, used to decompose long-horizon motion into a sequence of local goals. Navigation stacks chain waypoints through global and local planners, while in manipulation, waypoint-based policies predict sparse end-effector poses that a motion planner or low-level controller interpolates, trading control-rate reasoning for temporal abstraction. Why it matters for physical AI: Sparse waypoints compress long-horizon behavior into a few decisions, which shortens the effective horizon for learned policies and makes high-level plans from language models directly executable by conventional controllers. ### WebSocket Category: Software & Middleware URL: https://almond.bot/glossary/websocket WebSocket is a full-duplex communication protocol over a single TCP connection, standardized as RFC 6455, that enables low-latency bidirectional messaging between clients and servers after an HTTP upgrade handshake. Robotics systems use it for browser-based dashboards, remote monitoring, and teleoperation front ends, most notably through rosbridge, which exposes ROS topics, services, and actions to web clients as JSON over WebSocket. Why it matters for physical AI: Web-native transport lets operators supervise and teleoperate robots from any browser without installing middleware, which is how most fleet dashboards and remote intervention interfaces for deployed robots are built. ### Weed Control Robot Category: Applications URL: https://almond.bot/glossary/weed-control-robot A weed control robot is an agricultural robot that detects and eliminates weeds using mechanical tools, targeted micro-spraying, or high-power lasers, guided by plant-level computer vision that discriminates crops from weeds. Commercial systems such as Carbon Robotics' LaserWeeder operate row by row, and precision targeting can cut herbicide use dramatically relative to broadcast spraying. Why it matters for physical AI: Plant-by-plant weeding is a canonical perception-to-action task at field scale, demanding fine-grained detection, precise actuation, and all-weather reliability, and its economics are already competitive with chemical weed control. ### Weight Initialization Category: Robot Learning URL: https://almond.bot/glossary/weight-initialization Weight initialization is the choice of initial parameter values for a neural network before training, which governs early signal propagation and gradient scale. Schemes such as Xavier (Glorot) and Kaiming (He) initialization set layer variances to keep activations stable in deep networks. In robot learning the term also covers warm-starting: initializing a policy from pretrained vision or vision-language weights rather than from random values. Why it matters for physical AI: Initialization from pretrained foundation-model weights, rather than from scratch, is the single biggest lever behind sample-efficient robot policies, letting small demonstration datasets fine-tune capabilities learned from web-scale corpora. ### Welding Robot Category: Applications URL: https://almond.bot/glossary/welding-robot A welding robot is an industrial manipulator configured for arc, spot, or laser welding, combining a welding power source, torch or gun end effector, and coordinated motion control, often with external positioner axes. Welding was among the first large-scale robot applications, dominating automotive body shops since the 1980s; modern cells add seam tracking via laser scanners and through-arc sensing to compensate for part variation. Why it matters for physical AI: A worldwide shortage of skilled welders is driving demand for systems that program themselves from CAD or scan data, making welding a leading commercial target for perception-driven, learning-based automation of high-mix fabrication. ### Wheel Odometry Category: Navigation & SLAM URL: https://almond.bot/glossary/wheel-odometry Wheel odometry is the estimation of a mobile robot's pose change by integrating wheel rotations measured with encoders through the platform's kinematic model. It is cheap and high-rate but accumulates unbounded drift from wheel slip, uneven terrain, and tire-diameter error, so it is typically fused with IMU, visual, or LiDAR measurements in a filter rather than trusted alone. Why it matters for physical AI: Odometry remains the backbone prior for indoor mobile robot localization, and understanding its failure modes, especially slip, is essential when fusing it with learned or visual estimators on real fleets. ### Wheeled-Legged Robot Category: Locomotion URL: https://almond.bot/glossary/wheeled-legged-robot A wheeled-legged robot is a hybrid platform with actuated wheels mounted at the ends of articulated legs, combining the energy efficiency and speed of rolling on flat ground with legged capability for steps, curbs, and rough terrain. Examples include ETH Zurich's ANYmal on wheels and the Swiss-Mile robot, balancing bipeds such as Ascento, and wheeled variants of commercial quadrupeds. Why it matters for physical AI: Hybrid morphologies expand the deployable envelope of mobile robots in cities and warehouses, and their whole-body controllers, often trained with reinforcement learning, showcase how learned control exploits morphology beyond hand-designed gaits. ### Whisker Sensor Category: Perception URL: https://almond.bot/glossary/whisker-sensor A whisker sensor is a bio-inspired tactile sensor consisting of a slender, flexible shaft whose base deflections are measured to detect and localize contact, estimate surface shape, and sense airflow, emulating rodent vibrissae. Arrays of whiskers provide sparse but robust near-field geometry in conditions where cameras fail, such as darkness, dust, smoke, or optically ambiguous surfaces like glass. Why it matters for physical AI: Pre-contact and light-contact sensing fills the blind spot between vision and fingertip touch, giving manipulators and inspection robots collision-safe exploration in cluttered or visually degraded environments. ### Whole-Body Control Category: Control URL: https://almond.bot/glossary/whole-body-control Whole-body control (WBC) is a control framework that coordinates all of a robot's joints simultaneously to satisfy multiple prioritized objectives, such as maintaining balance, tracking an end-effector pose, and respecting contact and torque constraints. It is commonly formulated as a quadratic program over joint accelerations and contact forces solved at high rate, with strict or weighted task hierarchies arbitrating conflicts. Why it matters for physical AI: Humanoids and mobile manipulators must balance, reach, and avoid limits at once, and whole-body controllers are the standard layer that turns task-space commands, increasingly emitted by learned policies, into dynamically consistent joint torques. ### Wireless Robot Control Category: Teleoperation URL: https://almond.bot/glossary/wireless-robot-control Wireless robot control is the command and supervision of robots over radio links such as Wi-Fi, 5G, or proprietary RF, in place of tethered connections. System design must budget for latency, jitter, packet loss, and dropouts: teleoperation loops degrade noticeably beyond roughly 100 ms of round-trip delay, so deployments layer local safety behaviors, command smoothing, and autonomous fallback on top of the radio link. Why it matters for physical AI: Untethered operation is a precondition for mobile robots in homes, warehouses, and outdoor sites, and link quality directly bounds how much control can stay remote versus run onboard, shaping edge-compute architectures. ### Workcell Category: Applications URL: https://almond.bot/glossary/workcell A workcell is a self-contained arrangement of one or more robots together with the fixtures, part feeders, sensors, tooling, and safety equipment needed to perform a specific manufacturing task. Cell design fixes part presentation and robot reach, historically allowing blind, repeatable playback programs; modern cells add 3D vision and force sensing to tolerate part variation and reduce fixturing cost. Why it matters for physical AI: The economics of automation are decided at the cell level, and the shift from rigidly fixtured cells to perception-driven ones is precisely where learned manipulation earns its way into factories. ### Workspace Category: Math & Kinematics URL: https://almond.bot/glossary/workspace A workspace is the set of poses a robot's end effector can attain, determined by link lengths, joint limits, and kinematic structure. The reachable workspace contains every attainable position, while the dexterous workspace is the subset reachable with arbitrary end-effector orientation. Workspace boundaries coincide with kinematic singularities, near which dexterity and achievable velocity degrade. Why it matters for physical AI: Task feasibility starts with geometry: policies, planners, and cell layouts all inherit hard limits from the workspace, and mobile manipulation exists largely to relocate a finite arm workspace through a large environment. ### Workspace Analysis Category: Math & Kinematics URL: https://almond.bot/glossary/workspace-analysis Workspace analysis is the characterization of a manipulator's attainable pose set, including its boundary, volume, and dexterity distribution, computed analytically for simple chains or numerically by sampling forward kinematics into reachability maps. Quality measures such as Yoshikawa's manipulability index quantify how well-conditioned motion is at each configuration, guiding robot selection, base placement, and task positioning within a cell. Why it matters for physical AI: Precomputed reachability and dexterity maps let mobile manipulators choose base poses from which grasps are actually feasible, a quietly critical component in most manipulation-in-the-wild systems. ### World Frame Category: Math & Kinematics URL: https://almond.bot/glossary/world-frame A world frame is the fixed inertial coordinate frame in which robot poses, object locations, and map geometry are expressed, serving as the root of the transform tree that relates all other frames such as base, camera, and end-effector frames. Conventions matter in practice: ROS uses right-handed frames with z up, and transforms between frames are maintained by libraries like tf2. Why it matters for physical AI: Almost every multi-sensor bug traces back to frame confusion; consistent world-frame bookkeeping is what lets perception outputs, maps, and policy actions expressed in different frames compose into correct physical motion. ### World Model Category: Foundation Models URL: https://almond.bot/glossary/world-model A world model is a learned model of environment dynamics that predicts future observations or latent states conditioned on actions, enabling an agent to plan or train policies in imagination rather than solely through real interaction. Ha and Schmidhuber's 2018 work named the paradigm, the Dreamer line trains policies inside recurrent latent dynamics models, and large video-generation models are now positioned as general-purpose world simulators. Why it matters for physical AI: Real robot interaction is slow and costly, so learning inside a predictive model is a leading route to sample-efficient physical skill acquisition and to policies that anticipate consequences before acting. ### Worm Gear Category: Hardware URL: https://almond.bot/glossary/worm-gear A worm gear is a transmission that meshes a screw-like worm with a toothed wheel on a perpendicular axis, achieving high reduction ratios in a single compact stage. Sliding contact makes it less efficient than involute gearing, and at high ratios it self-locks, meaning the output cannot backdrive the input, so the joint holds position without motor power or brakes. Why it matters for physical AI: Self-locking joints hold loads unpowered, which benefits gravity-loaded axes and fail-safe designs, but the resulting non-backdrivability blocks torque transparency, making worm-geared joints poorly suited to compliant, contact-rich learned control. ### Wound Care Robot Category: Applications URL: https://almond.bot/glossary/wound-care-robot A wound care robot is a medical robotic system that assists in assessing and treating wounds, using 3D imaging and multispectral sensing to measure wound geometry and tissue state, and in research settings performing interventions such as debridement or dressing changes under clinician supervision. Most systems remain at the assessment-and-documentation stage, with autonomous physical treatment an active research area. Why it matters for physical AI: Chronic wound care is repetitive, protocol-driven, and facing acute nursing shortages, making it a representative target for safe physical AI in healthcare, where soft-tissue contact demands exceptional force sensitivity and reliability. ### Wrist Camera Category: Hardware URL: https://almond.bot/glossary/wrist-camera A wrist camera is a camera mounted on or near a manipulator's wrist, moving with the end effector to provide close-range, occlusion-resistant views of the object being manipulated; the configuration is also called eye-in-hand, in contrast to fixed eye-to-hand cameras. Because the viewpoint is a known function of the kinematics, wrist images simplify visual servoing and provide egocentric observations for learned policies. Why it matters for physical AI: Egocentric wrist views generalize across scenes better than fixed external views and capture the fine detail needed at contact, which is why nearly all modern manipulation data setups pair wrist and overview cameras. ### Wrist Configuration Category: Hardware URL: https://almond.bot/glossary/wrist-configuration A wrist configuration is the arrangement of the distal joints of a manipulator that orient the end effector, most commonly a three-axis wrist. A spherical wrist, whose three revolute axes intersect at a single point, decouples position from orientation and admits closed-form inverse kinematics by Pieper's condition; offset (non-spherical) wrists avoid the mechanical complexity of intersecting axes but generally require numerical IK. Why it matters for physical AI: Wrist geometry sets how a robot approaches grasps in clutter, where its orientation singularities fall, and whether inverse kinematics is fast and exact, all of which surface as hard constraints on manipulation policies. ## Glossary: X ### X-Ray Inspection Robot Category: Applications URL: https://almond.bot/glossary/x-ray-inspection-robot An X-ray inspection robot is a robotic system that positions radiographic sources and detectors to perform non-destructive testing of welds, castings, pipelines, and composite structures, often as a synchronized pair of manipulators moving source and detector on opposite sides of the part. Robotic computed tomography extends the approach, coordinating trajectories to reconstruct full 3D volumes of large components such as aerospace assemblies. Why it matters for physical AI: Automated non-destructive evaluation pairs precise coordinated motion with imaging physics, and it exemplifies inspection tasks where robots take over dull, hazardous work while machine learning increasingly automates the defect-detection step. ### XACRO (XML Macros) Category: Software & Middleware URL: https://almond.bot/glossary/xacro XACRO (XML Macros) is an XML templating language used in the ROS ecosystem to generate URDF robot descriptions, adding macros, parameters, conditionals, math expressions, and file inclusion so large robot models can be composed from reusable, parameterized components. A xacro file is expanded into plain URDF at build or launch time, keeping descriptions of arms, sensors, and grippers modular and maintainable. Why it matters for physical AI: Accurate, maintainable robot descriptions feed everything downstream, including kinematics, collision checking, simulation, and sim-to-real pipelines, and xacro is the de facto tool that keeps those descriptions consistent across robot variants. ### xArm (UFACTORY xArm) Category: Hardware URL: https://almond.bot/glossary/xarm The xArm is a family of low-cost collaborative manipulators from UFACTORY, offered in 5-, 6-, and 7-DOF variants with payloads of roughly 3 to 6 kg, integrated harmonic-drive joints, and Python, ROS, and Modbus interfaces. Its price point, an order of magnitude below traditional cobots at launch, made it a common platform in manipulation research labs and appears in several open imitation-learning and teleoperation stacks. Why it matters for physical AI: Cheap, ROS-friendly arms determine who can do embodied AI research; platforms like the xArm broadened participation and supplied heterogeneous embodiments to cross-robot datasets used for training generalist policies. ### Xavier (NVIDIA Jetson AGX Xavier) Category: Hardware URL: https://almond.bot/glossary/jetson-xavier Xavier (NVIDIA Jetson AGX Xavier) is an embedded system-on-module for edge AI, released in 2018, that pairs an 8-core ARM CPU with a 512-core Volta GPU and dedicated deep learning accelerators, delivering up to 32 TOPS within a 10 to 30 W power envelope. It became a standard onboard computer for drones, quadrupeds, and mobile manipulators before being superseded by the Jetson Orin generation. Why it matters for physical AI: Onboard inference budgets dictate which policies can run untethered, and the Jetson line defined the compute class around which a generation of autonomous robots, and the quantization tricks to fit models onto them, was designed. ### Xsens Motion Capture Category: Teleoperation URL: https://almond.bot/glossary/xsens-motion-capture Xsens motion capture is a wearable inertial motion-capture system, now sold under the Movella brand, that estimates full-body human kinematics by fusing measurements from body-worn IMUs through a biomechanical model, requiring no cameras or studio infrastructure. Unlike optical systems such as Vicon or OptiTrack, it works outdoors and in occluded spaces, at the cost of positional drift that must be constrained by fusion or contact assumptions. Why it matters for physical AI: Portable whole-body capture supplies the human motion driving humanoid teleoperation and retargeted imitation data, letting researchers record natural manipulation and locomotion anywhere rather than only inside instrumented capture volumes. ## Glossary: Y ### Yaskawa Motoman Category: Hardware URL: https://almond.bot/glossary/yaskawa-motoman Yaskawa Motoman is the industrial-robot business of Japan's Yaskawa Electric, one of the world's largest robot manufacturers alongside FANUC, ABB, and KUKA, with an installed base spanning arc welding, handling, assembly, and painting. Its lineup ranges from compact 6-axis arms to dual-arm and gantry systems, controlled by the MOTOMAN controller family, with ROS-Industrial drivers connecting them to research software stacks. Why it matters for physical AI: Incumbent industrial platforms represent the installed capacity that physical AI must integrate with, and vendor ecosystems like Motoman's define the reliability, safety certification, and uptime bar that learned manipulation has to meet. ### Yaw Category: Math & Kinematics URL: https://almond.bot/glossary/yaw Yaw is rotation about the vertical (z) axis of a body frame, the third component of the roll-pitch-yaw convention for describing 3D orientation; for a mobile robot or vehicle it is the heading angle. Yaw estimates typically come from fusing gyroscope integration with magnetometer, visual, or GNSS references, since gyro-only heading drifts and Euler-angle representations suffer gimbal lock at extreme pitch. Why it matters for physical AI: Heading error corrupts every downstream position estimate faster than translational noise does, which makes yaw observability a central concern in odometry, drone control, and any policy consuming orientation inputs. ### YCB Object Set Category: Data & Benchmarks URL: https://almond.bot/glossary/ycb-object-set The YCB object set is a standardized collection of 77 everyday physical objects, from cans and boxes to tools and toys, assembled by Yale, CMU, and UC Berkeley in 2015 so that manipulation research groups can benchmark on identical items. Each object ships with high-quality scanned mesh models, and derived benchmarks such as YCB-Video anchor 6-DOF pose estimation evaluation, including the BOP challenge. Why it matters for physical AI: Reproducibility in manipulation requires shared physical objects, not just shared code, and YCB remains the common vocabulary linking grasping papers, pose-estimation benchmarks, and simulation assets across a decade of research. ### YOLO (You Only Look Once) Category: Perception URL: https://almond.bot/glossary/yolo YOLO (You Only Look Once) is a family of real-time single-stage object detectors, introduced by Redmon et al. in 2016, that predict bounding boxes and class probabilities in one forward pass over the image instead of using a separate region-proposal stage. Successive versions from many maintainers improved accuracy while keeping inference fast enough for video rates on embedded GPUs, making YOLO the default detector in robotic vision pipelines. Why it matters for physical AI: Perception has to fit the control loop's latency budget on onboard compute, and YOLO-class detectors are what make closed-loop object-aware behavior feasible on drones, mobile robots, and pick-and-place cells. ### YuMi (ABB YuMi) Category: Hardware URL: https://almond.bot/glossary/yumi YuMi (ABB IRB 14000) is a dual-arm collaborative robot introduced by ABB in 2015, featuring two 7-DOF arms with approximately 500 g payload each, padded lightweight construction, and integrated grippers with optional vision, designed for small-parts assembly alongside human workers without safety fencing. It was among the first commercial dual-arm cobots and became a fixture in bimanual manipulation research. Why it matters for physical AI: Inherently safe dual-arm hardware anticipated today's push toward humanoid-style bimanual manipulation, and platforms like YuMi supplied early evidence for what coordinated two-arm learned policies require in sensing and compliance. ## Glossary: Z ### Z1 Arm (Unitree Z1) Category: Hardware URL: https://almond.bot/glossary/z1-arm The Z1 arm is a lightweight 6-DOF manipulator from Unitree Robotics with a payload of a few kilograms, integrated joint motors with torque control, and a weight low enough to mount on Unitree's quadrupeds. It is used both as a standalone low-cost research arm and as the manipulation payload in legged mobile manipulation, where arm and base are increasingly controlled by a single whole-body policy. Why it matters for physical AI: Affordable torque-controllable arms that ride on legged bases opened up loco-manipulation research, where reinforcement-learned whole-body controllers coordinate locomotion and reaching as one problem rather than two stacked systems. ### Zarr (Data Format) Category: Data & Benchmarks URL: https://almond.bot/glossary/zarr Zarr is an open-source storage format for chunked, compressed N-dimensional arrays, designed for efficient parallel reads and writes on local disks and cloud object stores. Robot learning pipelines use it to store large observation-action datasets because random access to chunks supports shuffled training without loading whole files; the Diffusion Policy reference implementation, among others, distributes demonstration data as Zarr archives. Why it matters for physical AI: Training generalist policies means streaming terabytes of multimodal trajectories into GPUs, and chunked array formats like Zarr are part of the data infrastructure that makes large-scale robot learning workloads practical. ### ZED Camera Category: Hardware URL: https://almond.bot/glossary/zed-camera The ZED camera is a family of stereo depth cameras from Stereolabs that compute depth from passive binocular stereo using GPU-accelerated, increasingly neural, matching rather than projected infrared patterns. This gives it longer range and reliable outdoor operation in sunlight, where active IR sensors such as the RealSense degrade, and the SDK adds visual-inertial odometry, spatial mapping, and human skeleton tracking. Why it matters for physical AI: Outdoor-capable depth at range is a prerequisite for field robots and sidewalk-scale autonomy, and integrated camera-plus-SDK packages shorten the path from raw sensing to the point clouds and odometry that navigation stacks consume. ### Zenoh Category: Software & Middleware URL: https://almond.bot/glossary/zenoh Zenoh is a high-performance publish-subscribe and query protocol from the Eclipse Foundation, developed by ZettaScale, that unifies data in motion and data at rest under one API and routes efficiently across LANs, WANs, and constrained links. It was adopted as an alternative ROS 2 middleware through rmw_zenoh, addressing long-standing DDS pain points such as discovery storms and unreliable behavior over Wi-Fi. Why it matters for physical AI: Fleet-scale robotics needs transport that spans robot, edge, and cloud without brittle network tuning, and Zenoh's rise inside ROS 2 reflects how deployment realities are reshaping core robot software infrastructure. ### Zero Dynamics Category: Control URL: https://almond.bot/glossary/zero-dynamics Zero dynamics are the internal dynamics that remain in a nonlinear system when its outputs are held identically at zero by feedback, revealing behavior that output control cannot see; systems with unstable zero dynamics are nonminimum-phase and fundamentally harder to control. In legged robotics, the hybrid zero dynamics (HZD) framework of Westervelt and Grizzle designs walking gaits as stable periodic orbits of this reduced dynamics. Why it matters for physical AI: Underactuated machines like bipeds cannot command every degree of freedom, so what happens in the unactuated remainder decides stability, informing both formal gait design and why some learned locomotion behaviors are fragile. ### Zero Moment Point (ZMP) Category: Locomotion URL: https://almond.bot/glossary/zero-moment-point The zero moment point (ZMP) is the point on the ground where the horizontal moment of the ground reaction forces vanishes; keeping it strictly inside the support polygon ensures a walking robot's feet do not tip. Formalized by Vukobratovic in the late 1960s, ZMP-based pattern generation, often via preview control of a linear inverted pendulum model, underpinned classic humanoids such as Honda's ASIMO and the HRP series. Why it matters for physical AI: ZMP remains the baseline notion of dynamic balance against which capture-point, MPC, and reinforcement-learned humanoid controllers are understood, and it still governs conservative walking modes in deployed bipeds. ### Zero-Shot Generalization Category: Robot Learning URL: https://almond.bot/glossary/zero-shot-generalization Zero-shot generalization is a model's ability to perform correctly on tasks, objects, instructions, or scenes never encountered during training, without any additional data or fine-tuning. In robot learning it is measured by evaluations on held-out object categories, novel spatial arrangements, unseen language commands, and new backgrounds, with vision-language-action models inheriting much of their zero-shot semantic breadth from web-pretrained backbones. Why it matters for physical AI: Homes and workplaces are open worlds that no dataset can enumerate, so zero-shot performance on the long tail, rather than in-distribution success rate, is the metric that decides whether robots are actually deployable. ### Zero-Shot Transfer Category: Robot Learning URL: https://almond.bot/glossary/zero-shot-transfer Zero-shot transfer is the deployment of a trained model in a new domain, embodiment, or environment without any adaptation data, most prominently sim-to-real transfer in which policies trained entirely in simulation run directly on hardware. Domain randomization over physics and appearance is the standard enabler, with quadruped locomotion policies among the clearest successes; cross-embodiment transfer between different robots is a newer frontier. Why it matters for physical AI: Every hour of real-robot fine-tuning is expensive and risky, so the degree to which simulation-trained or cross-robot policies transfer untouched sets the economics of scaling physical AI. ### Zero-Torque Mode Category: Control URL: https://almond.bot/glossary/zero-torque-mode Zero-torque mode is an operating mode in which a robot's actuators are commanded to exert no torque beyond, in most implementations, compensation for gravity and friction, so a person can move the arm freely by hand. It is the basis of kinesthetic teaching, where demonstrations are recorded by physically guiding the robot, and requires torque-controlled or backdrivable joints to feel genuinely transparent. Why it matters for physical AI: Hand-guiding is one of the most natural demonstration interfaces for manipulation data collection, and the quality of a robot's zero-torque behavior directly shapes how smooth and human-like the recorded trajectories are. ### Zeroth-Order Optimization Category: Robot Learning URL: https://almond.bot/glossary/zeroth-order-optimization Zeroth-order optimization is optimization using only function evaluations, without analytic gradients, estimating search directions from sampled perturbations as in evolution strategies, CMA-ES, and random search. It suits robotics problems where gradients are unavailable or unreliable, including hardware-in-the-loop tuning, controllers evaluated through contact-rich simulation, and policy search, where OpenAI's evolution strategies work showed it can rival gradient-based reinforcement learning. Why it matters for physical AI: Contact discontinuities and black-box simulators break backpropagation, so gradient-free methods remain a dependable fallback for tuning gaits, reward weights, and sim-to-real parameters directly against task outcomes. ## Glossary: # ### 1X NEO Category: Hardware URL: https://almond.bot/glossary/1x-neo 1X NEO is a humanoid robot developed by 1X Technologies for household environments, distinguished by quiet tendon-driven actuation, low-inertia limbs, and a soft textile covering intended to make close human contact safe. NEO is operated through a mix of learned autonomy and human teleoperation, with remote operators handling tasks the policies cannot yet complete while their interventions generate new training data; consumer preorders opened in 2025. Why it matters for physical AI: Home humanoids stress-test the full physical AI stack, including safe compliant hardware, data flywheels built on teleoperation, and honest boundaries between autonomy and remote assistance under real consumer expectations. ### 2-Finger Gripper Category: Hardware URL: https://almond.bot/glossary/2-finger-gripper A 2-finger gripper is an end effector with two opposing fingers, most commonly a parallel-jaw design in which both fingers translate along a single axis, actuated electrically or pneumatically. Devices like the Robotiq 2F-85 and the Franka gripper made it the default research end effector: a single grasp degree of freedom simplifies control and learning while handling a surprisingly broad range of objects, especially with compliant fingertips. Why it matters for physical AI: The vast majority of manipulation datasets and benchmark results assume a parallel jaw, so the two-finger gripper effectively defines the action space, and the capability ceiling, of most current learned manipulation. ### 3D Diffusion Policy Category: Robot Learning URL: https://almond.bot/glossary/3d-diffusion-policy 3D Diffusion Policy (DP3) is a visuomotor imitation learning method, introduced by Ze et al. in 2024, that conditions a diffusion-based action head on compact 3D representations extracted from sparse point clouds rather than on 2D images. The 3D scene encoding yields strong sample efficiency, learning many manipulation tasks from a few dozen demonstrations, and improved robustness to viewpoint and appearance variation. Why it matters for physical AI: Demonstrations are the scarcest resource in imitation learning, and results like DP3 argue that explicit 3D structure, not just bigger 2D backbones, is a lever for cutting data requirements per skill. ### 3D Printing (Additive Manufacturing) Category: Hardware URL: https://almond.bot/glossary/3d-printing 3D printing (additive manufacturing) is the fabrication of parts by depositing or fusing material layer by layer from a digital model, spanning processes from fused deposition modeling (FDM) in thermoplastics to selective laser sintering and metal powder-bed fusion. Robotics labs and companies rely on it for rapid iteration of grippers, fingertips, sensor mounts, and linkages, and research printers produce compliant mechanisms and soft robots directly. Why it matters for physical AI: Same-day iteration on end effectors and fixtures collapses the hardware-software co-design loop, letting teams adapt grippers to the failure modes their learned policies actually exhibit rather than designing around them in software. ### 3D Reconstruction Category: Perception URL: https://almond.bot/glossary/3d-reconstruction 3D reconstruction is the recovery of the three-dimensional geometry, and often appearance, of scenes or objects from sensor data such as multiple images, depth maps, or LiDAR scans. Classical pipelines combine structure-from-motion, multi-view stereo, and TSDF fusion into meshes, while neural approaches such as NeRF and 3D Gaussian splatting represent scenes as optimizable radiance fields with photorealistic novel-view synthesis. Why it matters for physical AI: Reconstructed geometry feeds grasp planning, collision checking, and digital twins, and fast photorealistic reconstruction increasingly closes the real-to-sim loop, turning captures of real workspaces into training environments for policies. ### 4-Bar Linkage Category: Hardware URL: https://almond.bot/glossary/4-bar-linkage A 4-bar linkage is a closed-chain mechanism of four rigid links joined by four revolute joints with one link fixed as ground, the simplest one-degree-of-freedom closed kinematic chain. By choosing link lengths, designers shape the coupler curve and mechanical advantage, which is why four-bars appear in gripper jaws, robot leg designs that keep actuator mass proximal, prosthetic knees, and countless latching and toggle mechanisms. Why it matters for physical AI: Clever passive kinematics can replace actuators, sensors, and control effort outright; linkage design is a reminder that in physical AI, intelligence embedded in mechanism trades directly against intelligence required in software. ### 5G Teleoperation Category: Teleoperation URL: https://almond.bot/glossary/5g-teleoperation 5G teleoperation is remote robot operation over 5G cellular networks, exploiting their higher bandwidth and lower radio latency relative to 4G, with the URLLC service class targeting latencies in the low milliseconds for control traffic. Deployments span remotely supervised sidewalk and delivery robots, teleoperated vehicles, and port and mining equipment, with network slicing and edge computing used to keep the control loop short. Why it matters for physical AI: Wide-area low-latency links make one remote operator per many robots economically viable, and human-in-the-loop fallback over cellular is the practical bridge deployment strategy while full autonomy matures. ### 6-Axis Force-Torque Sensor Category: Hardware URL: https://almond.bot/glossary/6-axis-force-torque-sensor A 6-axis force-torque sensor is a transducer, typically mounted between a robot's wrist and its end effector, that measures the full contact wrench: three orthogonal force components and three torque components. Most designs read strain gauges or capacitive elements on an elastic structure and require bias compensation for the payload's gravity. The measurements drive impedance and admittance control, insertion and polishing tasks, and collision detection. Why it matters for physical AI: Contact-rich assembly lives or dies on wrench feedback, and force-torque signals are increasingly recorded alongside images in demonstration datasets so learned policies can acquire force-aware, compliant behavior. ### 6-DOF Pose Estimation Category: Perception URL: https://almond.bot/glossary/6-dof-pose-estimation 6-DOF pose estimation is the task of recovering an object's full rigid-body pose, three translational and three rotational degrees of freedom, relative to a camera or world frame from visual or depth data. Methods span classical point-pair features on point clouds, learned approaches such as PoseCNN, and model-based or few-shot systems like FoundationPose, evaluated on benchmarks including YCB-Video and the BOP challenge suite. Why it matters for physical AI: Precise object pose is what converts detections into graspable geometry for model-based manipulation, and pose-aware pipelines still outperform end-to-end policies on high-precision tasks like tight-tolerance assembly. ### 7-DOF Arm Category: Hardware URL: https://almond.bot/glossary/7-dof-arm A 7-DOF arm is a manipulator with seven independent joints, one more than the six required to place an end effector at an arbitrary pose, making it kinematically redundant. The extra degree of freedom, often arranged in a human-like shoulder-elbow-wrist (S-R-S) structure as in the KUKA iiwa and Franka Emika Panda, lets the elbow swivel through a null space to avoid obstacles, joint limits, and singularities while holding the hand fixed. Why it matters for physical AI: Redundancy provides the workspace flexibility that cluttered human environments demand, and null-space motion gives whole-body and learned controllers room to optimize secondary objectives such as manipulability and collision clearance.