AI that senses the real world and moves a body in it, like a robot or self-driving car.
Physical AI senses its surroundings with cameras and other sensors, decides what to do, and then actually moves a body, such as a robot arm or wheels. Robots and self-driving cars are the best-known examples.
In 2023 Google DeepMind's RT-2 linked knowledge from web images and text directly to robot-arm control. Since collecting real-world practice data can be slow, costly or unsafe, developers also use models such as NVIDIA's Cosmos, released in 2025, to create virtual practice scenes.
In June 2026 BMW said the humanoid robot Figure 02 had helped build more than 30,000 X3 cars over about 11 months. But the release was joint with the robot's maker and doesn't say the figures were independently audited.
Think of it like this
Is a friend who's brilliant at football on a games console automatically good on a real pitch? The ball on screen always follows the same rules, but a real ball changes with the grass, the wind and how hard you kick. Physical AI is about fitting a mind trained on text and pictures to a real body on a real pitch.
For example
A robot that understands "pick up the red cup on the table and put it in the sink" and moves its arm to do it is physical AI.
Common mix-ups
Mix-upPhysical AI only means human-shaped humanoid robots.
ActuallyNVIDIA counts warehouse robots, robot arms, surgical robots and self-driving cars such as robotaxis as physical AI too. What matters is not the shape but whether a machine senses and acts in the real world on its own.
Mix-upPlug a smart chatbot AI into a robot and it will move like a person straight away.
ActuallyBalance has to be kept very fast. Separately from the part that understands meaning, Figure's Helix 02 runs fast loops that adjust whole-body balance 1,000 times a second and the joints 200 times a second. Thinking and moving work at different speeds.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- When a robot moves a cup, which parts are the AI's job and which are the machine's?
- Activity
- Split "pick up the cup and put it in the sink" into four boxes: see → understand → decide → move, and note what could go wrong in each. Read the RT-2 (E0250) and Gemini Robotics (E0532) records, find what the robot learned and mark which box it belongs in, then check who announced the numbers in the BMW record (E0983).
Records about this
Source-checked moments from AI history.
Figure's performance figures appeared on a BMW-branded channel for the first time. However, it is a joint release, and nothing in it says BMW independently audited those figures
Connected web-scale semantic knowledge to robot control as an output modality
Productized synthetic-data and simulation infrastructure for training physical AI
Transferred web-scale VLM knowledge directly to robot control, establishing the very category of VLA