The Physical AI Dictionary

Plain-English definitions for the words that show up in every humanoid robotics press release, paper and funding announcement — without assuming you already know them.

Humanoid robotics borrows vocabulary from three fields that barely talked to each other ten years ago: mechanical engineering, control theory, and machine learning. The result is a language where “policy” doesn’t mean rules, “embodiment” isn’t philosophy, and half the acronyms were coined in the last eighteen months.

This is a working glossary. Entries are grouped by theme rather than alphabetically, because most of these terms only make sense next to their neighbors — but there’s a search box and an A–Z index below if you arrived looking for one specific word.

96 terms · Last updated September 2026

Jump to a term (A–Z)

AAction chunking · Actuator · Affordance

BBackdrivability · Bimanual

CCenter of mass (CoM) · Cobot (collaborative robot) · Compliance · Configuration space (C-space) · Control frequency · Cross-embodiment · Cycle time (takt time)

DData flywheel · Demonstration · Dexterous hand · Diffusion policy · Digital twin · Direct-drive · Distributional shift (compounding error) · DoF (degrees of freedom) · Domain randomization · Dynamics

EEgocentric video · Embodied AI · Embodiment · Embodiment gap · Encoder · End effector · Episode (trajectory) · Exoskeleton and puppeteering rigs

FFall recovery · Fine-tuning · Flow matching · Force/torque sensor · Forward kinematics · Functional safety

GGait · Geared drive (harmonic, cycloidal, planetary) · Generalist policy (robot foundation model) · Generalization · Gripper

HHuman-in-the-loop · Humanoid

IImitation learning (behavior cloning) · IMU (inertial measurement unit) · Intervention rate (autonomy rate) · Inverse kinematics (IK)

JJoint · Joint space

KKinematics

LLatency · LeRobot · Link · Locomotion · Long-horizon task

MMassively parallel simulation · Mobile manipulation · Morphology · Motion capture (mocap) · MPC (model predictive control)

OOnboard compute (edge compute) · Open X-Embodiment

PPayload · PD controller · Physical AI · Physics engine · Policy · Proprioception

QQuasi-direct drive (QDD)

RRaaS (robots as a service) · Real-to-sim · Reinforcement learning (RL) · Retargeting · Reward function

SSample efficiency · Series elastic actuator (SEA) · Sim-to-real · Sim-to-real gap (reality gap) · Simulation · SLAM (simultaneous localization and mapping) · Success rate · Synthetic data · System identification

TTactile sensing · Task space (or Cartesian space) · Teleoperation · Tendon-driven · Trajectory · Transfer learning

UUMI (Universal Manipulation Interface) · Uptime and MTBF

VVLA (vision-language-action model)

WWhole-body control · World model

ZZero-shot · ZMP (zero moment point)

The big picture

The framing terms — what the field calls itself, and what it is betting on.

Physical AI: AI that perceives and acts in the real world through a body — sensors in, motor commands out. The industry term, popularized around 2024, and largely a rebrand of embodied AI for the era of foundation models and commercial humanoids. Used when talking to investors and executives.
Embodied AI: The older academic name for the same idea, in use since the 1990s. Used in papers and conference submissions. If someone switches between the two terms mid-sentence, they mean the same thing.
Morphology: The shape of a robot's body: how many limbs, legs or wheels, a pinching gripper or a suction cup, where the cameras sit. Morphology determines what the machine can physically do before any software is involved.
Humanoid: A robot whose body roughly approximates the human form — two legs, two arms, torso, head. The argument for the form factor is not aesthetics but infrastructure: stairs, doorways, handles and workbenches were all built for human bodies, so a human-shaped machine can in principle work in them without anyone rebuilding the building.
Generalist policy (robot foundation model): One large pre-trained model intended to control many robots across many tasks, then adapted to specifics — as opposed to hand-building a separate system per task. The central bet of the current wave of companies.

Hardware and mechanics

The body: what moves, what senses, and the trade-offs baked into each choice.

Actuator: The muscle. Any component that converts energy — usually electrical, sometimes hydraulic or pneumatic — into motion. In a modern humanoid this almost always means an electric motor plus a gearbox and its controller, packaged as one unit.
Joint: A movable connection between two rigid parts. Revolute joints rotate; prismatic joints slide. Nearly all humanoid joints are revolute.
Link: The rigid segments between joints — the bones.
DoF (degrees of freedom): The number of independent ways a robot can move. Each simple joint contributes one. A human arm is roughly seven; a modern humanoid is often 30–50 or more once you count the hands. More DoF means more capability and more of everything hard: weight, cost, wiring, control complexity.
Direct-drive: The motor is connected straight to the joint with no gearbox. Smooth, fast, and easy to sense forces through, because nothing sits between the motor and the world. The catch is that you need a large, heavy motor to produce useful torque.
Quasi-direct drive (QDD): The practical compromise: a motor with a low-ratio gearbox, typically under 10:1. Keeps most of the responsiveness and backdrivability of direct-drive while getting usable torque. Standard in modern legged robots.
Geared drive (harmonic, cycloidal, planetary): High-ratio gearboxes that turn fast, weak motor rotation into slow, strong joint rotation. Precise and powerful, but heavy, expensive, and very hard to push backwards — which matters for safety.
Tendon-driven: The motors sit away from the joint and pull cables routed to it, the way your forearm muscles pull on your fingers. This keeps mass out of the extremity, which is why most dexterous hands are tendon-driven. The costs are friction, cable stretch, wear, and control that is harder to get precise.
Backdrivability: Can you grab the limb and push it, and the motor turns? A backdrivable joint yields when it hits something unexpected, which makes it much safer around people and much better at tasks involving contact. A heavily geared joint resists and will happily keep pushing.
Compliance: Springiness — deliberate give in the system, either mechanical or created in software. A compliant robot absorbs a collision instead of transmitting it.
Series elastic actuator (SEA): An actuator with a spring intentionally placed between the motor and the joint. Measuring how far the spring deflects gives you the force for free, and the spring absorbs impacts.
End effector: Whatever is on the end of the arm: gripper, hand, welding torch, suction cup.
Gripper: The simple end effector — usually two parallel fingers or a vacuum cup. Reliable, cheap, limited to objects it was designed around.
Dexterous hand: A multi-fingered hand aiming at human-like manipulation. Brutally hard engineering: many degrees of freedom in a small volume, with actuation, wiring and tactile sensing all competing for the same space.
Payload: How much weight the robot can lift or carry, usually quoted per arm and often measured in the most flattering possible pose.
Proprioception: The robot's sense of its own body — joint angles, velocities and torques — as opposed to exteroception, which is sensing the outside world through cameras and lidar.
Encoder: The sensor that measures a joint's angle. Every controlled joint has at least one.
IMU (inertial measurement unit): Accelerometer plus gyroscope. Tells the robot which way is down and how fast it is rotating. Balance is impossible without it.
Force/torque sensor: Measures forces where the limb meets the world, usually at the wrist or ankle. What lets a robot push a drawer closed firmly and set an egg down gently.
Tactile sensing: Touch. Contact, pressure, slip and texture at the fingertips or across a skin. Widely agreed to be a bottleneck for dexterous manipulation, and still not solved at reasonable cost or durability.

Motion and control

How a machine works out where to put its limbs, and stays upright while doing it.

Kinematics: The geometry of motion: how joint angles relate to where the hand ends up. Ignores forces entirely.
Forward kinematics: Given the joint angles, where is the hand? Straightforward arithmetic.
Inverse kinematics (IK): Given where you want the hand, what joint angles get you there? Much harder. There are often infinitely many answers, sometimes none, and the useful ones avoid hitting the robot's own body.
Dynamics: The forces side: what torques are actually required to produce a motion, accounting for inertia, gravity and momentum.
Joint space: Describing motion as a list of joint angles — the language the hardware natively speaks. "Shoulder to 30°, elbow to 95°."
Task space (or Cartesian space): Describing motion as the position and orientation of the hand in 3D space — the language the task natively speaks. "Hand here, palm down." A large fraction of robot control is translating between task space and joint space.
Configuration space (C-space): The abstract space of every pose the robot could adopt. Motion planning happens here.
Trajectory: A path through space with timing attached — not just where to go, but when to be there.
Whole-body control: Coordinating every joint at once so that reaching a high shelf also involves bending the knees, shifting the hips and adjusting the arms for balance. Fundamentally different from controlling an arm bolted to a table, where the base cannot fall over.
Locomotion: Getting around: walking, running, stairs, uneven ground.
Gait: The repeating pattern of a walk — which foot lifts when, and for how long.
ZMP (zero moment point): A classic balance criterion: keep this computed point inside the area covered by the feet and the robot won't tip. Underpinned two decades of walking controllers, and is steadily being displaced by learned ones that don't reason in these terms at all.
Center of mass (CoM): The single point that behaves like the average of all the robot's mass. Most balance reasoning is about where it sits relative to the feet.
Fall recovery: Getting up unaided after falling over. Unglamorous, rarely demoed, and a hard commercial requirement — a machine that needs two people and a crane after every fall is not deployable.
MPC (model predictive control): Control by repeatedly simulating a few seconds into the future, choosing the best plan, executing a fraction of it, then throwing it away and re-planning. Computationally expensive, very effective, still the backbone of a lot of locomotion.
PD controller: The simplest useful feedback loop: push proportionally to how far off you are, damp proportionally to how fast you're moving. Sits underneath nearly every more sophisticated method on the stack.
Control frequency: How often commands are updated. Low-level joint control typically runs at 500–1000 Hz; a large neural policy might only produce actions at 10–50 Hz. Bridging that gap is a real engineering problem — see action chunking.
Latency: The delay between sensing something and acting on it. Small numbers matter enormously when the robot is falling over.
Bimanual: Two-armed. Coordinating two arms on the same object is substantially harder than twice one arm, because each arm constrains the other through the object.
Mobile manipulation: Moving and manipulating at the same time, rather than driving somewhere, stopping, then reaching. Hard because the base moving changes what the arms can reach and destabilizes them while they work.
SLAM (simultaneous localization and mapping): Building a map of an unknown space while simultaneously working out where you are within it.
Affordance: What an object offers for action — a handle affords grasping, a surface affords placing. A common way of framing what perception systems need to extract.

How robots learn

The model architectures and training methods behind the current wave.

Policy: The robot's decision function: observations in, actions out. When people say a model was "trained," the policy is the thing that was trained.
Imitation learning (behavior cloning): Training a policy to copy human demonstrations. Conceptually simple and effective, but data-hungry and fragile in situations the demonstrations never covered.
Reinforcement learning (RL): Learning by trial and error against a reward signal. Dominant for locomotion, where you can crash a million times in simulation for free. Much harder for manipulation, where realistic contact is hard to simulate and real-world trials are slow and expensive.
Reward function: How behavior is scored in RL. Writing one that produces the behavior you actually wanted — rather than a robot that exploits a loophole and vibrates in place forever — is a discipline of its own.
Distributional shift (compounding error): The classic imitation-learning failure. A small deviation puts the robot in a state slightly unlike anything in its training data, so its next action is slightly worse, which puts it further off, and so on until it's doing something absurd. Explains a lot of failed demos.
VLA (vision-language-action model): A model that takes camera images and a natural-language instruction and outputs robot actions. Built by extending vision-language models — the architecture behind image-understanding chatbots — with a third capability: producing physical motion. The dominant architecture of the current era, and the reason robotics suddenly moved on foundation-model timelines.
World model: A learned simulator. A model trained to predict what the world will look like after a given action, so an agent can imagine consequences before committing to them — and so developers can generate synthetic data at scale. Distinct from a physics engine in that nobody programmed the physics; it was inferred from video.
Diffusion policy: A policy that generates actions using the same denoising process as image generators. Its advantage is handling multimodality: when there are several equally valid ways to do a task, simpler methods average them into a useless compromise, while a diffusion policy can commit to one.
Flow matching: A related generative technique, generally faster to sample from than diffusion. Several recent VLAs use it in place of diffusion for exactly that reason.
Action chunking: Predicting a short sequence of future actions in one go — typically 8 to 16 steps — and executing several of them before re-planning. Produces smoother motion, suppresses jitter, and lets a slow model drive a fast robot.
Long-horizon task: A task with many dependent steps, like making coffee or unloading a dishwasher. Hard because errors compound, and because the robot has to track how far through it is.
Zero-shot: Doing something never seen in training — a new object, a new room, a new phrasing of the instruction. The word appears in a lot of marketing; ask what specifically was novel.
Fine-tuning: Taking a large pre-trained model and training it further on a small, task-specific dataset. The standard way of getting from a generalist policy to something that works in your warehouse.
Sample efficiency: How much data or experience is needed to reach a given level of performance. The central economic question in robot learning, because real robot data is slow and expensive to collect, unlike text scraped from the web.
Generalization: Working outside the exact conditions of training: different lighting, a different mug, someone else's kitchen. The gap between an impressive demo and a product.

Where the data comes from

Robot data is not scraped from the web. These are the ways it gets collected.

Demonstration: One recorded example of a task being performed correctly. The basic unit of imitation-learning data.
Episode (trajectory): One complete recorded run — the full time-series of observations and actions from start to finish. What a dataset is counted in.
Teleoperation: A human directly controlling the robot in real time, through VR controllers, an exoskeleton, motion capture, or a small twin of the robot arm moved by hand. Serves two quite different purposes: collecting training data, and acting as the fallback when autonomy fails in a deployment.
Intervention rate (autonomy rate): How often a human has to take over. The honest measure of how good a deployed system is, and conspicuously absent from most demo videos.
Motion capture (mocap): Recording human movement precisely using markers and camera arrays, or instrumented suits. A major source of full-body motion data for humanoids, since the whole point of a human-shaped robot is that human movement should transfer to it.
Egocentric video: First-person footage from a head-mounted camera, roughly matching what a robot's head camera would see. Attractive because ordinary people can record enormous quantities of it while doing ordinary tasks. The limitation is that it shows what hands did, not what forces they applied or what joint angles produced them.
UMI (Universal Manipulation Interface): A handheld gripper with a camera mounted on it that a person uses to perform tasks directly — no robot involved. Because the tool's geometry matches a robot gripper, the recorded motion maps onto the robot afterwards. The appeal is cost and portability: data can be collected anywhere, by anyone, without a robot present.
Exoskeleton and puppeteering rigs: A wearable frame that records joint angles as a person moves, or a scaled-down replica of the robot arm that the operator moves by hand while the real arm mirrors it.
Retargeting: Translating motion recorded on one body onto a different one — human motion capture onto a robot with different limb proportions and joint limits, or data collected on robot A onto robot B. Almost never lossless: a human shoulder has a range of motion no current robot shoulder can reproduce, so something has to be approximated or thrown away.
Human-in-the-loop: A workflow where a person watches and corrects the robot during operation, with those corrections fed back as additional training data.
Data flywheel: The loop everyone is trying to start: deployed robots generate data, data improves the model, a better model raises autonomy, higher autonomy justifies more deployment.
Open X-Embodiment: A large open dataset pooling demonstrations from dozens of labs across many different robot types — the landmark attempt at shared robot data, and the basis for a lot of cross-embodiment research.
LeRobot: Hugging Face's open-source robot learning library and dataset hub. Effectively the field's shared plumbing for smaller teams.

Simulation

Training in a physics engine, and everything that goes wrong on the way out of it.

Simulation: Running the robot inside a physics engine instead of reality. Faster than real time, perfectly repeatable, and nothing breaks or needs recharging.
Physics engine: The software that does the simulating — MuJoCo, Isaac Sim, Bullet and others. They differ mainly in how they handle contact, which is exactly the part robotics cares about most.
Massively parallel simulation: Running thousands of copies of the robot simultaneously on a GPU. This is the development that made modern reinforcement-learning locomotion practical: years of walking experience in an afternoon.
Sim-to-real: Training a policy in simulation and then deploying it on physical hardware.
Sim-to-real gap (reality gap): Everything the simulator gets wrong: friction, contact and deformation, motor delay, gear backlash, cable stretch, sensor noise, the fact that the floor is slightly slippery today. The reason a policy that looks perfect in simulation can fall over in the first two seconds of reality.
Domain randomization: Deliberately varying simulation parameters during training — friction, masses, lighting, latency, sensor noise — so the policy can't rely on any one of them being exact. The aim is that reality ends up looking like just another variation it already handles.
System identification: Measuring the real robot's actual physical parameters and feeding them back into the simulator so the two match more closely. The complement to domain randomization: one narrows the gap, the other makes the gap not matter.
Real-to-sim: The reverse direction — building a simulation from captured reality, such as scanning a real kitchen to train in a digital copy of it.
Digital twin: A simulated replica of a specific real machine or facility, kept in sync with its physical counterpart.
Synthetic data: Training data that was generated rather than recorded — from simulation, or increasingly from world models and video generators.

Bodies and transfer

Why data collected on one body does not simply run on another.

Embodiment: The specific body a policy is running on: its morphology, joint layout, limb lengths, sensors and limits.
Embodiment gap: The mismatch between the body that produced the data and the body that has to execute it. Human hands have five fingers, tendons and skin; a robot gripper has two rigid pincers. This is why you cannot simply pour human video into a robot and get behavior out, and why retargeting exists.
Cross-embodiment: Training one policy on data from many different robot bodies, so that it learns the structure of tasks rather than the quirks of one machine. The hoped-for payoff is that data collected on any robot makes every robot better — which would change the economics of the whole field.
Transfer learning: Reusing what was learned in one setting to accelerate learning in another.

Deployment and business

The numbers that decide whether a robot is bought, and the rules that let it work.

Success rate: The percentage of attempts completed. Always check the denominator, the task definition and whether a human was standing nearby.
Cycle time (takt time): How long one repetition takes, and the rate a production line requires. This is the number that decides whether a robot is economically worth installing, regardless of how impressive it looks.
Uptime and MTBF: How much of the working day the machine is actually productive, and the mean time between failures. Humanoids have many joints, and every joint is a thing that can fail.
RaaS (robots as a service): Leasing the robot by the hour or month rather than selling it outright. The prevailing humanoid business model, since it turns a large capital purchase into an operating cost comparable to a wage.
Onboard compute (edge compute): The processing done on the robot itself rather than offloaded to a server. Constrained by power, heat and weight — and required for anything where a network round-trip would be too slow.
Cobot (collaborative robot): An industrial arm rated to operate near people without a safety cage, usually by limiting speed and force.
Functional safety: The certification regime governing whether a machine is allowed near humans — standards like ISO 10218 and ISO/TS 15066. Deeply unglamorous, and one of the real barriers between a viral demo and a deployment.