Understand
Physical AI glossary
The essential terms, defined in English and French.
Action tokens Tokens d'action models
Discrete symbols that encode robot actions, joint deltas, gripper commands, end-effector poses, so a transformer can predict them exactly as a language model predicts words. Tokenizing actions lets robotics reuse the entire LLM toolchain: the same architecture reads an instruction, looks at images, and emits action tokens that are decoded into motor commands.
Actuator Actionneur hardware
The component that converts energy into motion: typically an electric motor plus gearbox in modern robots, sometimes hydraulic or pneumatic in legacy or high-force systems. Actuator characteristics, torque density, backdrivability, bandwidth, efficiency, define what a robot can do physically. Humanoid economics hinge on actuators, which represent a large share of the bill of materials.
AI Act AI Act (règlement européen sur l'IA) regulation
EU Regulation 2024/1689, the first comprehensive law governing artificial intelligence, applying a risk-based approach: banned practices, high-risk systems with strict obligations, and lighter transparency rules elsewhere. AI used as a safety component of machinery is classified high-risk. Obligations phase in through 2026-2027, and the Act works alongside CE marking and the Machinery Regulation for robots.
Automated Guided Vehicle (AGV) AGV (véhicule à guidage automatique) fundamentals
A driverless transport vehicle that follows fixed guidance infrastructure: buried wires, magnetic tape, reflectors, or QR codes on the floor. AGVs are reliable and predictable but stop when their path is blocked and require infrastructure changes to modify routes. They remain widespread in mature intralogistics installations, increasingly displaced or complemented by more flexible AMRs.
Autonomous Mobile Robot (AMR) Robot mobile autonome (AMR) fundamentals
A mobile robot that navigates dynamically using onboard sensors, SLAM-based maps, and path planning, rather than following fixed guidance. AMRs reroute around people and obstacles in real time, which makes them suited to shared, changing environments such as warehouses and hospitals. They are the flexible successor to AGVs for intralogistics and material transport.
Autonomy Autonomie fundamentals
The degree to which a robot performs tasks without human intervention, ranging from teleoperation through shared control and supervised autonomy to full autonomy. In industrial deployments, autonomy is always bounded: an operational design domain defines the environments and tasks within which the system is validated to operate, with human takeover or safe stop outside those limits.
Behavior cloning Behavior cloning (clonage de comportement) models
The simplest form of imitation learning: supervised training of a policy to reproduce the actions a demonstrator took in each observed state, treating demonstrations as a labeled dataset. It is easy to implement and scales with data, but suffers from distribution shift: small errors take the robot into states absent from training data, where predictions degrade further.
Benchmark simulation
A standardized task suite and metric set used to compare robot learning methods fairly: simulation suites such as RLBench, LIBERO, or ManiSkill, and physical protocols with fixed objects and success criteria. Robotics benchmarks are harder to standardize than language ones because hardware, resets, and environments vary, which is why reported success rates deserve scrutiny.
Brownfield / greenfield industry
Two deployment contexts: greenfield means a new site designed from scratch, where automation can shape the layout; brownfield means an existing facility with legacy equipment, tight spaces, and running production that must not stop. Most real-world automation is brownfield, which favors flexible robots, AMRs, cobots, humanoids, that adapt to the environment rather than requiring its redesign.
CE marking Marquage CE regulation
The manufacturer's declaration that a product meets all applicable EU requirements, mandatory before placing machinery on the European market. For robot systems it requires a risk assessment, compliance with relevant directives and regulations, harmonized standards such as ISO 10218, a technical file, and a declaration of conformity. Integrators typically CE-mark the complete robot cell, not just the robot.
Collaborative robot (cobot) Robot collaboratif (cobot) fundamentals
An industrial robot arm designed to work alongside humans without safety fencing, using power and force limiting, speed monitoring, and rounded, sensorized designs. Cobots trade payload and speed for safety and ease of programming, making automation accessible to small manufacturers. Universal Robots popularized the category; safety requirements are specified in ISO 10218 and ISO/TS 15066.
Cross-embodiment models
Training a single model on data collected from many different robot bodies, arms, mobile manipulators, humanoids, so skills transfer across hardware instead of being locked to one platform. The Open X-Embodiment dataset, aggregating demonstrations from dozens of labs and robot types, showed that cross-embodiment training improves performance even on robots seen only sparsely in the data.
Degrees of freedom (DoF) Degrés de liberté (DoF) fundamentals
The number of independent motions a mechanism can perform, typically one per actuated joint. A standard industrial arm has six DoF, enough to place a tool at any position and orientation in its workspace. Humanoids commonly exceed thirty DoF, and dexterous hands add twenty or more, which multiplies control complexity and cost.
Depth camera Caméra de profondeur hardware
A camera that captures per-pixel distance in addition to color, using stereo vision, structured light, or time-of-flight measurement. Depth cameras such as Intel RealSense or the ZED series are the workhorse 3D sensor for manipulation: they feed grasp planning, obstacle avoidance, and the observation space of learned policies at a fraction of LiDAR cost.
Dexterous hand Main dextre hardware
A multi-fingered robotic hand with many actuated degrees of freedom, often sixteen to twenty-plus, designed to approach human manipulation ability: in-hand reorientation, tool use, handling varied objects. Dexterous hands combine miniaturized actuation, tendon or linkage transmission, and increasingly tactile sensing. They remain expensive and fragile, which is why many humanoids ship with simpler grippers first.
Diffusion policy models
A manipulation policy that generates action sequences with a diffusion model, progressively denoising random noise into a coherent trajectory conditioned on observations. Popularized in 2023, the approach handles multimodal solutions gracefully: when several valid ways to grasp exist, it commits to one instead of averaging them. Diffusion heads are now common inside VLA architectures.
Digital twin Jumeau numérique simulation
A virtual replica of a physical asset, cell, or entire factory, kept synchronized with its real counterpart through sensor and process data. Digital twins let engineers validate robot programs, simulate throughput, train AI policies, and rehearse layout changes before touching the shop floor. Platforms such as NVIDIA Omniverse and Siemens Xcelerator anchor this market.
Domain randomization simulation
A sim2real technique that randomizes simulation parameters during training, friction, masses, motor strength, textures, lighting, camera pose, so the learned policy becomes robust to variation. If the policy succeeds across thousands of randomized worlds, the real world looks like just another sample. It is the key ingredient behind most successful simulation-trained locomotion and manipulation controllers.
Edge computing hardware
Running computation close to where data is produced, on the robot or on a nearby industrial PC, rather than in a remote cloud. For physical AI, edge execution is usually mandatory: control loops cannot tolerate network latency or outages, and factory data often cannot leave the site. Platforms such as NVIDIA Jetson embody this approach.
Embodied AI IA incarnée fundamentals
A research field studying agents that learn through a body, physical or simulated, interacting with an environment. Its core hypothesis is that intelligence emerges from the sensorimotor loop between perception and action, not from passive data alone. Embodied AI underpins modern robot learning, from autonomous navigation to dexterous manipulation, and is often used interchangeably with physical AI.
End effector Effecteur terminal fundamentals
The tool mounted at the end of a robot arm that interacts with the workpiece: a two-finger gripper, suction cup array, welding torch, screwdriver, or multi-fingered hand. End effector choice often determines what tasks a cell can perform and drives a large share of integration effort. Quick-change systems let one robot swap tools automatically.
Fine-tuning models
Continuing the training of a pre-trained model on a smaller, task-specific dataset so it specializes without learning from scratch. In physical AI, a generalist VLA is typically fine-tuned on a few hundred demonstrations of the target task, robot, and workcell. This is how integrators adapt foundation models to a customer's parts, fixtures, and lighting conditions.
FMEA (Failure Mode and Effects Analysis) AMDEC industry
A systematic method for identifying how a product or process can fail, rating each failure mode by severity, occurrence, and detectability, and prioritizing corrective actions accordingly. FMEA is mandatory practice in automotive and standard in robot cell design. AI-driven systems complicate it: failure modes of learned policies are harder to enumerate than those of deterministic mechanisms.
Force-torque sensor Capteur force-couple hardware
A sensor, usually mounted at a robot's wrist, that measures forces and torques along all six axes. It enables force-controlled operations such as polishing, assembly with tight tolerances, and safe contact detection. Many modern cobots achieve similar capability by estimating forces from motor currents in each joint, trading accuracy for lower cost.
Foundation model Modèle de fondation models
A large neural network pre-trained on broad, diverse data that can be adapted to many downstream tasks through fine-tuning or prompting, rather than being built for a single purpose. In robotics, foundation models aim to provide general skills, perception, language understanding, and manipulation priors, that transfer across robots and tasks, replacing per-application engineering.
Harmonic drive Réducteur harmonique (harmonic drive) hardware
A compact gear mechanism based on a flexible toothed cup, the flexspline, deformed by an elliptical wave generator, achieving very high reduction ratios with near-zero backlash in a small volume. Also called strain wave gearing, it dominates robot joints where precision matters. Its cost and the concentration of suppliers make it a strategic component in humanoid supply chains.
High-risk AI system Système d'IA à haut risque regulation
Under the AI Act, an AI system whose failure could harm health, safety, or fundamental rights, including AI acting as a safety component of regulated products such as machinery. High-risk status triggers obligations: risk management, data governance, technical documentation, logging, human oversight, robustness, and conformity assessment before market placement. Most industrial physical AI falls in this category.
Human oversight Supervision humaine regulation
The AI Act requirement (Article 14) that high-risk AI systems be designed so humans can effectively monitor them, understand their outputs, intervene, and stop them. For robots this translates into supervision interfaces, interpretable status reporting, override and emergency stop mechanisms, and operating models where autonomy stays within human-validated bounds rather than replacing accountability.
Humanoid robot Robot humanoïde fundamentals
A general-purpose robot with a human-like body plan, typically two legs or a wheeled base, two arms, and a torso, designed to operate in spaces built for people: factories, warehouses, eventually homes. The humanoid form factor promises task versatility without redesigning the workplace. Leading platforms include Tesla Optimus, Figure, Agility Digit, Unitree, and Boston Dynamics Atlas.
Imitation learning Apprentissage par imitation models
Training a robot policy from human demonstrations, usually collected by teleoperation or kinesthetic guidance, instead of hand-coding behaviors or defining reward functions. It is currently the dominant recipe for manipulation skills because demonstrations encode subtle contact strategies that are hard to specify. Its main limits are data cost and compounding errors when the robot drifts from demonstrated states.
IMU (Inertial Measurement Unit) Centrale inertielle (IMU) hardware
A sensor package combining accelerometers and gyroscopes, sometimes magnetometers, that measures linear acceleration and angular velocity at high frequency. IMUs are essential for balance control in legged robots and for odometry in mobile platforms, where they bridge the gaps between slower vision or LiDAR updates. Their drift over time requires fusion with other sensors.
Inference Inférence models
Running a trained model to produce outputs, as opposed to training it. For robots, inference has hard real-time constraints: a manipulation policy must produce actions every few tens of milliseconds, often on embedded hardware with limited power. Latency, memory footprint, and determinism drive choices such as quantization, distillation, and running smaller models at the edge.
Intralogistics Intralogistique industry
The flow of materials, goods, and information inside a site: receiving, storage, order picking, transport between workstations, and shipping. Intralogistics is the largest current market for mobile robots, with AMR fleets, automated storage systems, and picking arms orchestrated by warehouse software. It is where physical AI first reached industrial scale.
ISO 10218 regulation
The international safety standard for industrial robots, in two parts: 10218-1 covers the robot itself, 10218-2 covers the integrated robot system and cell. The 2025 revision integrates collaborative operation requirements previously found in ISO/TS 15066 and strengthens functional safety expectations. Compliance with ISO 10218 is the backbone of CE marking for robot installations.
ISO/TS 15066 regulation
The technical specification that defined how humans and robots may safely share a workspace, detailing the four collaborative modes and, critically, quantified pain-onset thresholds for force and pressure on different body regions under power and force limiting. Its content has largely been absorbed into the 2025 revision of ISO 10218, but the document still shapes cobot application design.
JEPA (Joint Embedding Predictive Architecture) JEPA (architecture prédictive à plongement joint) models
A self-supervised architecture, introduced by Yann LeCun in 2022, that predicts a future or masked observation in an abstract representation space rather than reconstructing raw pixels or tokens, using an energy-based model with separate context and target encoders. Implemented as I-JEPA (images, 2023), V-JEPA and V-JEPA 2 (video, 2024-2025), it underpins AMI Labs' world models and is positioned as an alternative to autoregressive LLMs for planning and reasoning about the physical world.
LiDAR hardware
A sensor that measures distance by timing laser pulses reflected off surfaces, producing precise 2D scans or 3D point clouds independent of ambient light. LiDAR anchors AMR navigation and safety functions: certified safety laser scanners define protective fields that slow or stop the robot when a person enters. Solid-state designs are cutting cost and size.
Machine tending Machine tending (chargement de machines) industry
Automating the loading and unloading of production machines, CNC mills, lathes, injection molders, presses, with a robot that places raw parts, retrieves finished ones, and may handle door opening or chip blowing. It is one of the highest-volume robot applications and a prime target for physical AI, since part variety traditionally demands costly re-engineering.
Machinery Regulation (EU 2023/1230) Règlement Machines (UE 2023/1230) regulation
The EU regulation replacing the 2006 Machinery Directive, applicable from January 2027, that sets essential health and safety requirements for machinery sold in Europe. It explicitly addresses new technologies: self-evolving AI behavior, safety functions driven by machine learning, collaborative applications, and cybersecurity affecting safety. Together with the AI Act, it forms the legal frame for industrial physical AI.
MTBF (Mean Time Between Failures) MTBF (temps moyen entre pannes) industry
The average operating time between failures of a repairable system, a core reliability metric in industrial procurement. Industrial robot arms routinely exceed 60,000 to 100,000 hours MTBF, a bar that humanoids and AI-driven systems are far from meeting today. Buyers use MTBF with MTTR, mean time to repair, to estimate availability and maintenance cost.
NPU (Neural Processing Unit) NPU (unité de traitement neuronal) hardware
A processor specialized for neural network operations, mainly the matrix multiplications behind inference, delivering far better performance per watt than a general-purpose CPU. In robots and edge devices, NPUs run perception and policy models within tight power budgets. They typically execute quantized models, INT8 or FP16, and their software toolchains matter as much as their raw specs.
OEE (Overall Equipment Effectiveness) TRS (taux de rendement synthétique) industry
The standard measure of manufacturing productivity, multiplying three ratios: availability (uptime versus planned time), performance (actual versus ideal speed), and quality (good parts versus total). World-class plants target 85 percent. Any robot or AI system pitched to a factory is ultimately judged on its effect on OEE, not on demo impressiveness.
Open weights Poids ouverts (open weights) models
A release model where the trained parameters of a neural network are published for anyone to download, run, and fine-tune, though training data and code may stay private. Open-weight robotics models such as OpenVLA and GR00T N1 let integrators and researchers build on state-of-the-art policies without API dependence, a key enabler for on-premise industrial deployments.
OpenUSD simulation
Universal Scene Description, an open framework originated at Pixar for describing, composing, and exchanging complex 3D scenes across tools. In physical AI, OpenUSD is becoming the common language of digital twins and simulation: robots, factories, and props described once can flow between CAD, Omniverse, Isaac Sim, and rendering pipelines without lossy conversion.
Palletizing Palettisation industry
Stacking cases, bags, or products onto pallets in stable, space-efficient patterns, and the reverse operation, depalletizing. It is a classic robot application driven by ergonomics and labor scarcity. Modern systems compute mixed-case stacking plans on the fly and use vision to depalletize unknown loads, a task long considered too variable for automation.
Perception-action loop Boucle perception-action fundamentals
The continuous cycle in which a robot senses its environment, interprets the data, decides on an action, executes it, and observes the result, feeding the next cycle. Loop frequency matters: locomotion controllers run at hundreds of hertz while high-level reasoning may run at a few hertz. Closing this loop robustly is the defining challenge of physical AI.
Photorealistic rendering Rendu photoréaliste simulation
Generating simulated camera images realistic enough, through ray tracing, physically based materials, and accurate lighting, that perception models trained on them work on real footage. Photorealism narrows the visual part of the reality gap and is central to synthetic data pipelines for detection, segmentation, and vision-based policies. It complements, rather than replaces, domain randomization.
Physical AI IA physique fundamentals
Artificial intelligence that perceives, reasons about, and acts on the real physical world through machines such as robots, vehicles, and industrial equipment. Unlike purely digital AI, it must cope with sensor noise, real-world physics, latency, and safety constraints. The term went mainstream in 2024-2025 as foundation models began controlling robots directly, from warehouse arms to humanoids.
Physics engine Moteur physique simulation
Software that computes rigid-body dynamics, contacts, friction, and constraints to simulate how objects and robots move. Robotics-grade engines, MuJoCo, PhysX, Isaac Sim's GPU pipeline, Genesis, prioritize contact accuracy and massive parallelism: thousands of environments run simultaneously on one GPU for reinforcement learning. Engine fidelity directly bounds how well simulation-trained skills transfer to hardware.
Pick-and-place industry
The task of grasping items from one location and placing them at another: order picking, kitting, sorting, packaging. Structured pick-and-place with known parts is solved; the frontier is unstructured picking of varied, jumbled, or deformable items from bins and totes, where learned vision and grasping policies now outperform classical engineered pipelines.
Policy Policy (politique de contrôle) models
In robot learning, the function that maps observations, such as camera images and proprioceptive state, to actions, such as joint targets or end-effector motions. A policy can be a small task-specific network or a large VLA. Training methods include imitation learning from demonstrations and reinforcement learning from reward signals; deployment requires the policy to run within real-time control budgets.
Proprioception fundamentals
A robot's sense of its own body state: joint positions and velocities, motor torques, orientation, and balance, measured by encoders, current sensing, and inertial units. Proprioceptive feedback runs at high frequency and is essential for locomotion, force control, and contact-rich manipulation. It complements exteroception, the perception of the external world through cameras and LiDAR.
Reality gap Reality gap (écart sim-réel) simulation
The mismatch between a simulator and the physical world: unmodeled friction and contact dynamics, sensor noise, latency, lighting, wear, and manufacturing variation. The reality gap is why policies that score perfectly in simulation can fail on hardware. Closing it drives most simulation research, through better physics, learned residual models, and randomization strategies.
Reinforcement learning (RL) Apprentissage par renforcement (RL) models
A learning paradigm where an agent improves its policy through trial and error, guided by a reward signal rather than labeled examples. In robotics, RL is typically run at massive scale in simulation, then transferred to hardware. It excels at dynamic skills such as legged locomotion and recovery behaviors, where robust controllers emerge from millions of simulated falls.
Robots-as-a-Service (RaaS) RaaS (Robots as a Service) industry
A business model where customers pay a subscription or per-task fee for robotic capability instead of buying hardware outright. The vendor retains ownership, handles maintenance and software updates, and carries performance risk. RaaS lowers the adoption barrier for SMEs and cash-constrained operations, and is common in intralogistics, cleaning, and increasingly humanoid pilot deployments.
Safety PLC Automate de sécurité (safety PLC) regulation
A programmable logic controller with redundant, self-diagnosing architecture certified to execute safety functions: emergency stops, light curtain monitoring, safe speed limits, zone control. Safety logic runs on this certified channel, deliberately separate from standard automation and from any AI. In AI-driven cells, the safety PLC is the deterministic layer that constrains what learned policies can physically do.
SIL / PL regulation
Two scales quantifying the reliability required of a safety function: Safety Integrity Level (SIL 1-3 for machinery, from IEC 61508/62061) and Performance Level (PL a-e, from ISO 13849). The required level follows from risk assessment: severity, exposure, avoidability. Typical collaborative robot safety functions must reach PL d, which mandates redundant architectures and proven failure rates.
Sim2real simulation
The process of transferring behaviors learned in simulation to physical robots. Simulation offers unlimited, cheap, safe training data, but simulators never match reality perfectly, so policies must be hardened through domain randomization, accurate system identification, and sometimes real-world fine-tuning. Sim2real is the standard pipeline for locomotion and increasingly credible for manipulation.
SLAM (Simultaneous Localization and Mapping) SLAM (localisation et cartographie simultanées) fundamentals
A family of algorithms that lets a robot build a map of an unknown environment while simultaneously estimating its own position within that map, using LiDAR, cameras, or both fused with inertial data. SLAM is the foundation of AMR navigation and is challenged by dynamic scenes, long corridors, and visually repetitive environments.
SoC (System on Chip) SoC (système sur puce) hardware
A single chip integrating CPU cores, GPU, memory controllers, and increasingly a dedicated neural accelerator, plus interfaces for cameras and sensors. SoCs are the compute heart of robots: they minimize size, power, and cost compared with separate components. Examples include NVIDIA's Jetson Orin and Thor modules and Qualcomm's robotics platforms.
Synthetic data Données synthétiques simulation
Training data generated by simulation or generative models rather than collected in the real world: rendered images with perfect labels, simulated trajectories, or video generated by world models. Synthetic data attacks robotics' central bottleneck, the scarcity of robot demonstrations, and provides rare or dangerous scenarios on demand. Its value depends on how well it matches real distributions.
System 1 / System 2 Système 1 / Système 2 models
An architecture pattern for robot control borrowed from cognitive science: a slow, deliberate System 2, usually a vision-language model reasoning at a few hertz, sets goals and decomposes tasks, while a fast System 1 policy executes reactive motor control at tens or hundreds of hertz. Figure's Helix and NVIDIA's GR00T N1 are prominent implementations.
System integrator Intégrateur industry
A company that turns robots and components into working production systems: process design, gripper and fixture engineering, safety compliance, PLC integration, installation, and support. Robot manufacturers rarely deploy directly; integrators own the customer relationship and carry project risk. For physical AI vendors, the integrator channel largely determines how fast new technology reaches factories.
Tactile sensor Capteur tactile hardware
A sensor that measures contact: pressure distribution, shear forces, slip, vibration, or texture at the robot's fingertips or skin. Technologies range from capacitive and piezoresistive arrays to vision-based skins such as GelSight, which image the deformation of a soft membrane. Tactile feedback is considered the missing ingredient for reliable manipulation of fragile, deformable, or occluded objects.
Takt time Temps takt industry
The pace at which products must be completed to match customer demand: available production time divided by required output. A takt of 60 seconds means one finished unit per minute. Takt time is the hard constraint any automation must satisfy; a robot that is flexible but slower than takt does not fit a production line.
TCO (Total Cost of Ownership) TCO (coût total de possession) industry
The full cost of an automation system over its life: purchase price plus integration, programming, training, energy, maintenance, spare parts, downtime, and eventual decommissioning. Integration often costs two to three times the robot itself. Physical AI's economic promise is to slash the integration and reprogramming share of TCO, not merely the hardware price.
Teleoperation Téléopération fundamentals
Remote control of a robot by a human operator, through joysticks, motion-capture suits, VR interfaces, or leader-follower arms. Beyond intervention in hazardous environments, teleoperation has become the main method for collecting demonstration data used to train manipulation policies by imitation learning. Many humanoid demos shown publicly mix autonomous behavior with teleoperated segments.
TOPS (Tera Operations Per Second) TOPS (téra-opérations par seconde) hardware
A headline metric for AI accelerator throughput: trillions of low-precision operations per second, usually quoted for INT8 arithmetic. TOPS enables rough comparisons between edge chips but is easy to overread: real performance depends on memory bandwidth, model architecture, software stack, and sustained thermal limits. Two chips with identical TOPS can differ several-fold on an actual robot workload.
Vision-Language Model (VLM) Modèle vision-langage (VLM) models
A foundation model trained jointly on images and text, able to describe scenes, answer questions about them, and ground language in visual content. In robotics, VLMs serve as the semantic backbone: they identify objects, interpret instructions, and plan task steps, while a separate control policy executes the motions. VLAs are built by extending VLMs with action outputs.
Vision-Language-Action model (VLA) Modèle VLA (vision-langage-action) models
A neural network that takes camera images and a natural-language instruction as input and outputs robot actions directly, extending vision-language models with an action head. VLAs generalize across objects and tasks better than task-specific controllers. Landmark models include RT-2, OpenVLA, pi-zero from Physical Intelligence, and NVIDIA GR00T, most trained on large multi-robot demonstration datasets.
World model World model (modèle de monde) models
A learned model that predicts how an environment will evolve, given current observations and candidate actions. World models let agents imagine outcomes before acting, plan in latent space, and train policies inside their own simulation. In physical AI they power video-prediction foundation models such as NVIDIA Cosmos and are considered a key route toward robots that reason about physics.