Physical AI

Physical AI means AI that directs robots, aircraft, lab instruments, delivery vans and warehouse stock, choosing each next action itself as the physical world changes. Most AI software only produces text, images or numbers on a screen. Physical AI acts on real objects, so a wrong decision can break something or hurt someone.

Here is how it works in general. Sensors and cameras observe the real surroundings. The AI reads the observations, decides what to do next, and sends a command to a machine, such as a robot arm, a vehicle or a pump. The machine acts, and the sensors observe the new situation. The cycle repeats many times a second. Because it can make mistakes, the system is given limits: places it must not go, forces it must not exceed, and a person who can stop it.

Limits come in several forms: a boundary the machine must not cross, a top speed, a maximum force, and an emergency stop. A person may approve each action, which is called human in the loop, or may only watch and step in, which is called human on the loop. The AI must also answer within the time that the physical task allows. A robot arm cannot wait two seconds to decide. Such systems are tested first in simulation. The real world is untidy, with poor light, slippery floors and objects the system has never seen, so rare cases are the main risk.

Take, for example, a warehouse robot that picks parcels. It looks at a pile, decides which parcel to lift, and adjusts when one slips; see robotic picking. Consider a drone that surveys a farm and changes its route when wind rises; see drone autonomy. Say a laboratory machine runs an experiment, reads the result, and chooses the next experiment itself; see self-driving labs.

This idea belongs to the group of pages on AI agents: AI agents, agent tools and limits and robot foundation models. How far a machine acts alone is described in levels of autonomy, and proof of safety in safety certification.

How is it different from an AI that works on a desk? An AI that drafts an email can be corrected before it is sent. Physical AI acts on the real world at once, so it needs limits and a way to stop it.