AI that has to act, not just answer
For most of the last decade, 'AI' mostly meant models that consume text or images and produce text or images back. A chatbot reads your question and writes an answer. An image classifier looks at a photo and outputs a label. None of that output has to survive contact with the physical world, if the model is wrong, nothing falls over and nothing breaks. Physical AI is the term the robotics industry has converged on for models that instead have to perceive the physical world and produce actions that move a real body through it, where being wrong has physical consequences like a dropped object, a collision, or a fall.
This distinction matters because it changes what 'understanding' has to mean. A language model can be extremely fluent about the concept of pouring water without ever having to get the wrist angle, timing, and grip force right. A physical AI system has to close that gap, its output is a stream of motor commands that either works in three-dimensional space with real gravity, friction, and momentum, or it doesn't. Physical AI is not a new kind of neural network architecture, it is a new job description for AI models, one that demands grounding in physics and real-time feedback that pure text and image models never had to deal with.