Physical AI is AI that perceives, reasons about, and acts in the physical world, typically through robots, vehicles, or machines. It contrasts with AI that stays in the digital realm, such as chatbots or document tools.
Physical AI combines perception (cameras, lidar, force sensors), a model that decides what to do, and actuation (motors, grippers, wheels). The term gained wide use in the 2020s, but it largely continues earlier work in robotics and embodied AI.
Current approaches rely heavily on simulation and large models. Robots are trained in simulated environments, often on synthetic data, then transferred to real hardware. Foundation models for robotics, including vision-language-action models, aim to let one model handle many tasks and robot types. Humanoid robots, AMRs, and learning-based manipulators are common examples.
The gap between demo and production is large. A robot that works in a lab must also work across shifts, lighting changes, part variation, and with safety certification. Physical AI systems act on the world, so errors can damage equipment or injure people. That makes validation, safety standards, and fallbacks more important than in purely digital AI. The term is also used heavily in marketing; ask what the system actually does.
Questions about Physical AI
Is physical AI new?
The term gained wide use in the 2020s, but it largely continues earlier work in robotics and embodied AI. It is also used heavily in marketing, so ask what the system actually does.
What is the difference between physical AI and AI such as chatbots?
Physical AI acts in the physical world through robots, vehicles or machines. Chatbots and document tools stay in the digital realm. Because physical AI acts on the world, its errors can damage equipment or injure people, so validation, safety standards and fallbacks matter more.
How are physical AI robots trained?
Current approaches rely heavily on simulation and large models. Robots are trained in simulated environments, often on synthetic data, then transferred to real hardware. The main limit of synthetic data is the domain gap: models trained only on synthetic data often perform worse on real data.
Physical AI in real plants
- Hyundai Motor Group and Boston Dynamics: Boston Dynamics opened a testbed and training center at Hyundai Motor Group Metaplant America where Atlas humanoid robots are trained on logistics and sequencing of automotive parts.
- BMW Group and Figure AI: In a pilot completed in 2025, Figure 02 humanoid robots inserted sheet-metal parts for welding in the body shop at BMW Group Plant Spartanburg.
- BMW Group and Figure AI: BMW Group will start using Figure 03 humanoid robots at Plant Spartanburg for a logistics sequencing use case: sorting delivered components into sequencing trolleys.
- Catalyst Brands and Figure AI: Catalyst Brands and Figure AI announced a commercial partnership starting at the Reno, Nevada Distribution Logistics Center, where humanoids will first support pouch sorting system sequencing.
- Toyota Motor Manufacturing Canada and Agility Robotics: After a year-long pilot, Toyota Motor Manufacturing Canada signed a Robots-as-a-Service agreement for seven Digit humanoids to load and unload parts totes on a tugger at its Woodstock plant.
Vendors linked to Physical AI
Agility Robotics (4) · Figure AI (3) · Boston Dynamics (1) · FANUC (1) · Physical Intelligence · Universal Robots
Physical AI in the digest
Related terms
Foundation models for robotics
Foundation models for robotics are large models pretrained on broad data, such as images, text, and many robots' demonstrations, that can be adapted to control different robots across many tasks, instead of training one narrow model per task.
Synthetic data
Synthetic data is artificially generated data that mimics some statistical properties of real data.
Cobot (collaborative robot)
A cobot (collaborative robot) is an industrial robot designed to work in close proximity to people, with safety functions that limit speed, force, or distance so that fencing can be reduced or removed after a risk assessment.
AMR (autonomous mobile robot)
An AMR (autonomous mobile robot) is a wheeled robot that navigates a facility using onboard sensors and maps, planning its own route and steering around obstacles.
Bin picking
Bin picking is a robotic task in which a robot, guided by a vision system, locates and grasps individual parts from a container where they lie in random positions and orientations.