What is physical AI?

Physical AI is AI with a body. Chatbots produce words and images; physical AI controls machines that move through the real world, like humanoid robots, warehouse robots and self-driving cars. It has to see its surroundings, understand what's going on and decide how to act, safely and in real time.

How it works

• Sensing. Cameras, lidar and other sensors feed the robot a constant picture of the world.
• Vision-language-action models. These models combine what the robot sees with instructions like "fold the towel" and output the movements to do it.
• World models. AI that learns how the physical world behaves, including gravity, motion and space, so it can predict what happens next. NVIDIA's Cosmos models are one example.
• Practice in simulation. Robots are trained in physics-based virtual worlds and with synthetic data, trying a skill thousands of times through trial and error before touching the real world.

Who's building it

• NVIDIA offers Isaac GR00T, an open platform for humanoid robots, and Cosmos world models for robots and autonomous vehicles.
• Google DeepMind launched Gemini Robotics in March 2025, and Gemini Robotics 1.5 in September 2025, which can reason before acting and transfer skills between different robots.
• Physical Intelligence, a startup, released its π0 model in 2024, trained on data from several robot types and shown folding laundry and assembling boxes.

Why it's hard

• Data is scarce. There's no internet-sized library of robot movements, so companies lean on simulation and synthetic data.
• Generalizing. A robot that works in one kitchen or on one arm often fails on another.
• Safety. Mistakes in the physical world can hurt people. Google DeepMind uses layered safety checks and released a dataset to test robot safety.

Physical AI is why robotics news is accelerating: the same advances behind chatbots are now being applied to machines that can act.