What Is Physical AI?

Physical AI is the software that lets a robot perceive its surroundings and take physical action. Here is what it means, how it works, and why it matters for humanoids.

Physical AI is the software that lets a machine sense the physical world and act in it. A chatbot predicts the next word. Physical AI predicts the next move: how hard to grip a cup, where to step, when to stop. It is the difference between a model that talks about folding laundry and a robot that actually folds it.

The current approach most humanoid companies are chasing is the vision-language-action model, or VLA. The robot takes in camera and sensor data, understands a goal in plain language, and outputs motor commands directly, without a programmer scripting each step. Figure’s Helix and Tesla’s work with Grok are both attempts at this. The promise is a robot you instruct in words that learns new tasks by watching rather than by reprogramming.

The hard part is data. Language models trained on the entire internet. There is no internet-scale archive of robots doing physical chores, so companies are generating that data themselves through teleoperation, simulation, and fleets of deployed robots recording what they do. This is why several home robots, including NEO, lean on remote human operators early on. Every guided task is also a training example.

Physical AI is the bottleneck for the whole category. The hardware is further along than the intelligence that drives it.