Models & learning
Physical AI
Physical AI is artificial intelligence used in machines that perceive and act in the physical world, including robots, autonomous vehicles, and other autonomous systems. It closes a loop between sensing, decisions, action, and physical consequences. The label is emerging rather than standardised and often overlaps with embodied AI.
Also known as: physical artificial intelligence
Updated
Intelligence with physical consequences
Physical AI describes AI systems whose outputs change the physical world and generate new sensor input. A robot may observe a scene through cameras, interpret a task, issue motor commands, make contact with an object, and then adjust after measuring the result. The physical feedback loop distinguishes the system from an AI model that only returns text, images, or labels.
The term covers more than humanoid robots. The CSET primer includes robots, autonomous vehicles, and smart spaces, while NVIDIA uses it for autonomous machines that perceive, understand, and perform actions in the real world. Physical AI is therefore a domain label, not one model architecture.
Relationship to embodied AI
Physical AI and embodied AI are often used interchangeably, but there is no settled industry definition. CSET notes that organisations also use phrases such as embodied machine intelligence and generative physical AI for overlapping parts of the AI-robotics field.
On Humanoids Data, physical AI refers specifically to systems acting through real hardware. Embodied AI is broader and can include agents acting inside persistent simulated environments. The distinction is useful because simulation can model physical consequences, while only deployment exposes the system to the target hardware, sensor noise, contact dynamics, damage, and safety constraints.
Why training data matters
A physical AI model needs data that connects observations to actions and outcomes. Web images or human video can provide visual and task knowledge, but they do not normally contain the motor commands, body state, contact measurements, or timing of a particular robot.
Real robot demonstrations provide that grounding. Simulation and world models can expand coverage, generate labels, and create rare scenarios, but their outputs still need validation against representative hardware. Our guide to physical AI training data explains how human video, robot demonstrations, synthetic data, tactile signals, and failure records contribute different parts of the learning problem.
What the label does not guarantee
Calling a system physical AI says nothing by itself about autonomy, reliability, generalisation, or safety. A learned model may handle visual variation while depending on a fixed worktop, hand, controller, and reset procedure. Conventional control, force limits, monitoring, and human intervention can remain necessary even when a foundation model chooses the task or action.
The International Federation of Robotics identifies reliability, cycle time, energy use, maintenance, dexterity, safety, and security as practical tests for humanoid deployment. Those measures are more informative than the physical AI label alone.
Sources
Related terms
Models & learning
Embodied AI
Embodied AI is artificial intelligence that perceives and acts through a body in an environment, so its actions change the observations available to it next. The body may be a physical robot or, in common research usage, an agent situated in a persistent simulated world. Embodied AI emphasises the coupled loop between morphology, perception, action, learning, and the environment.
Hardware & control
Humanoid robot
A humanoid robot is a robot whose body is modelled on the human form, typically with a torso, head and limbs, so it can move through human spaces or use human-scale tools and interfaces. The term describes morphology, not intelligence or autonomy; wheeled, partial-body and simplified-hand designs are also sometimes classed as humanoids, so the robot's actual configuration should be stated.
Models & learning
Robot foundation model
A robot foundation model is a broadly pretrained model intended to provide a reusable starting point for multiple robot tasks, environments or embodiments. It learns from diverse robotics and sometimes web or human data, then acts directly or is adapted with target-domain data. The term describes a training and reuse strategy, not one fixed architecture.
Models & learning
Vision-language-action model
A vision-language-action model is a robot policy that conditions on visual observations and natural-language instructions to predict executable actions. It adapts a pretrained vision-language model, or a similar multimodal backbone, using robot trajectories so that semantic knowledge is connected to an embodiment’s action space. The action representation may use discrete tokens or continuous values.
Models & learning
World model
A world model is a learned predictive model of how an environment changes, often conditioned on a robot’s current state and candidate actions. It may forecast pixels, latent states, rewards or other task-relevant quantities. A policy or planner can use those predictions to compare possible futures before commanding the physical robot.
Simulation & transfer
Digital twin
A digital twin is a fit-for-purpose digital representation of a specific physical robot, asset or process that is kept synchronised with its real counterpart through operational data. It may contain geometry, dynamics and simulation models, but the maintained link to an identified real system distinguishes it from an ordinary, standalone simulator.