What is Physical AI?

Physical AI is AI that leaves the screen: it no longer only generates text or images, it perceives a real environment, decides on an action and executes it through a body, whether a smart robot on a factory floor, a self-driving vehicle or an automated production line. It is the natural extension of agentic AI: an agent that calls an actuator instead of an API. The practical difference is not trivial: a software agent's mistake is undone with a rollback, a physical agent's mistake can damage machinery or put a person at risk, which makes the safety margins and controls far stricter than anything required of a chatbot or a digital assistant. The term spans a wide range, from a robotic arm on a factory floor to a self-driving vehicle on public roads, but the common thread stays the same: before acting, the system must correctly perceive its environment and predict the consequences of its own action with a much narrower margin for error than pure software tolerates.

Where the hype comes from (and what is real)

The term took off when NVIDIA CEO Jensen Huang called it, at CES 2026, the next "ChatGPT moment": a leap in capability for systems that act in the physical world, comparable to what language models represented for text. A terminology note is worth making: you will often hear "embodied AI" used as an almost interchangeable synonym. The difference is subtle, not substantial: Physical AI is the umbrella term for the whole stack (data, models, hardware), embodied AI emphasizes the perception and physical body-control component. In business practice, it is Physical AI that is gaining ground as the term of choice.

Before the robot comes the data

The point headlines about humanoid robots tend to skip: no robotic arm or autonomous vehicle works without a solid data foundation underneath. You need sensors producing reliable streams (IIoT), infrastructure that processes that data in real time close to the source (edge AI), a digital representation of the plant to simulate before acting (the digital twin) and, upstream of all that, models capable of predicting the physical consequences of an action (world models). Physical AI is not a robotics project: it is a data project that, once mature enough, also enables robotics. For an Italian manufacturing company the useful question today is not "which robot to buy" but "is my process data already ready to feed these systems once they mature". In most cases the honest answer is no, and that is where you start.