Most AI applications operate entirely in software. Physical AI closes the loop with the world: a model takes in camera or sensor input, works out what to do, and issues actions that move a physical system like a robot arm or a walking robot. The underlying capabilities are the same ones other AI systems need, including vision, language understanding, and tool use. What changes is that the model runs against real-time constraints, and its mistakes land in a room rather than a chat window. Two neighboring capabilities are spatial intelligence, reasoning about layout and motion, and world models, which predict how the environment will respond to an action.
The current examples connect commodity robotics hardware to general-purpose multimodal models, often called through an API rather than run on the robot, letting a robot follow spoken commands or navigate by sight without task-specific programming. A harder target, still mostly unbuilt, is automated scientific experimentation, where a model both designs an experiment and physically runs it.
“Done” is not the physical result
An illustrative task: a robot is asked to put a blue block in a tray. A model may identify the block and decide what to do, while motion planning and low-level control work out reachable positions and move the arm. Then a camera check afterward shows the block landed beside the tray. The model’s “done” was a claim about the plan, not a measurement of the world. Learned policies can merge some of these layers, so this is a division of responsibilities, not a rule that every step needs its own model. When you evaluate a physical AI system, score the observed outcome, and note where each part runs: software on a robot can still call a remote model, which puts network delay inside the control loop.
Sources
- NVIDIA Glossary: What is Physical AI? — Contrasts generative AI, which creates content, with physical AI, which helps autonomous machines perceive, reason about and act in the real world; checked October 26, 2026.
Go deeper
From the conversation
-
Gemini 3 & Robot Dogs: Inside Google DeepMind's AI Experiments | Paige Bailey -
Most of the Web Will Never Get APIs for AI Agents | Dhruv Batra, Yutori -
Agents Are Overtaking Humans in Search: Why Search Must Be Rebuilt -
How Cisco Secures Personal AI Assistants for 90,000 Employees | DJ Sampath -
Genspark's Bet: AI Agents Become the Users of Software | Wen Sang -
Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10