// DefinedTermSet

Physical AI glossary

Plain, citable definitions for the vocabulary of physical AI and RobotOps — the terms the rest of the reporting leans on, from the lab stack to the fleet floor.

01Embodied AI: AI that perceives and acts through a physical body — a robot, vehicle or manipulator — rather than only processing text or images on a screen. It must handle real-world physics, uncertainty and consequences that a purely digital model never faces.
02Vision-language-action (VLA) model: A foundation model trained to turn camera input and a language instruction directly into robot motor commands — the closest thing physical AI has to a general-purpose brain, in contrast to a narrow model built for one task.
03Sim-to-real transfer: Training a robot policy in simulation — where trial and error is cheap and safe — and then getting it to work on a physical robot. The gap between the two is one of the hardest open problems in robotics; it is why demo footage can mislead.
04Digital twin: A simulated replica of a robot, workcell or facility, kept in sync with sensor data from its physical counterpart. Teams use it to test changes, train models and predict failures before touching the real hardware.
05Teleoperation: Remote human control of a robot, often used today not to run production but to collect the demonstration data that trains an autonomous policy — a person in a haptic rig or VR headset generating the examples a model later learns from.
06World model: An internal representation an AI system builds of how its environment behaves and reacts to actions, letting it predict outcomes before acting instead of only reacting to what it sees. Physical AI leans on world models to plan multi-step tasks safely.
07RobotOps: The operational discipline of running a fleet of robots in production — telemetry, over-the-air updates, incident response, predictive maintenance and interoperability — modelled on DevOps, but for machines that can bump into the physical world.
08Fleet telemetry: The continuous stream of health and performance data — battery state, joint temperatures, task completion, near-misses — that every robot in a fleet reports back, normalised into one schema so operators can compare units across vendors.
09Over-the-air (OTA) update: Pushing new software, model weights or firmware to a robot fleet remotely rather than servicing each unit by hand. Done well it includes staged rollout and automatic rollback, since a bad update on a physical machine is riskier than on a phone.
10Predictive maintenance: Using telemetry patterns to flag a robot component likely to fail before it does, so it can be serviced on a schedule instead of causing an unplanned stoppage. It is one of the clearest RobotOps wins because downtime on a fleet compounds fast.
11Autonomous mobile robot (AMR): A robot that navigates a facility on its own, planning routes around obstacles and people, as opposed to an AGV that follows a fixed physical path such as a wire or magnetic strip. AMRs are the dominant form factor in today’s warehouse deployments.
12ROS 2: The open-source Robot Operating System (version 2), the middleware most robotics teams use to wire together sensors, planning and control code across many machines. It is not an operating system in the Windows/Linux sense — it runs on top of one.
13Collaborative-robot safety (ISO 10218 / 15066): The international standards defining how a robot may safely share space with people without a cage — limiting force, speed and pressure on contact, and requiring safety-rated monitored stops. Any physical AI deployment near workers has to clear this bar, AI Act aside.
14Robot unit economics: The full cost of running one robot for one hour or shift — hardware amortisation, energy, compute, maintenance and supervision — set against the value of the work it replaces or adds. It is the number that separates a fundable deployment from a funded demo.
15Humanoid robot: A robot built with a roughly human body plan — two legs, two arms, an upright torso — on the bet that a world built for human bodies (doorways, tools, vehicles, stairs) is easiest to operate in with a human-shaped machine, not the cheapest form factor for any single task.
16Actuator: The component that converts electrical, hydraulic or pneumatic energy into physical motion at a robot joint — the motor-plus-gearing assembly that actually moves the arm or leg. Actuator cost and quality dominate a humanoid robot’s bill of materials.
17Degrees of freedom (DoF): The number of independent ways a robot joint or body can move — a human hand alone has roughly 20-25. More degrees of freedom mean finer, more human-like manipulation, but each one adds an actuator, a failure point and a control problem to solve.
18Nvidia Isaac / GR00T: Nvidia’s robotics platform: Isaac Sim/Lab for simulation and synthetic-data generation, and GR00T, a family of foundation models for humanoid perception and control. Widely cited as reference infrastructure across the industry, though usage numbers are largely vendor-reported.
19CAGR (compound annual growth rate): The smoothed annual growth rate a market would need to go from a starting value to an ending value over a period, assuming steady compounding. In robotics market forecasts it is routinely quoted with wide swings between research firms, because the underlying market definition varies as much as the number itself.
20EU Machinery Regulation: The EU regulation (replacing the older Machinery Directive) that sets safety and conformity requirements for machines placed on the EU market, including robots — increasingly discussed alongside the EU AI Act because a robot with an AI-driven safety function can fall under both regimes at once.
21Edge compute: Processing that runs onboard a robot itself rather than in the cloud — necessary because a robot cannot wait out network latency to decide whether to stop before a collision. Edge compute boards like Nvidia Jetson trade raw power for the low latency and offline reliability physical AI needs.
22VDA 5050: An open interface standard that lets AMRs and AGVs from different manufacturers talk to the same fleet-management software, avoiding vendor lock-in when a facility runs mixed-brand robots side by side. Adoption is real but partial — plenty of fleets still run on proprietary protocols.
23Imitation learning (behavior cloning): Training a robot policy directly on recorded human demonstrations — usually teleoperation data — rather than having it discover behavior through trial and error. It is the dominant way VLA models learn today, precisely because real-world trial and error is slow, costly and sometimes unsafe.
24Robots-as-a-Service (RaaS): A subscription pricing model where a customer pays a monthly fee per robot instead of buying the hardware outright, shifting maintenance and upgrade risk to the vendor. It lowers the upfront capital barrier but changes the payback-period math — a RaaS quote and a capex quote are not directly comparable.