Physical AI, Chapter 1: Defining Physical AI

This is the first post in a series walking through my new book Physical AI: Engineering Embodied Intelligence for the Real World. Over the next ten days I will take one chapter a day. We start with the gap the whole book is about — the distance between knowing what an omelette is and being able to make one with a body.

A language model can write you a recipe for an omelette, in iambic pentameter if you ask, in a few hundred milliseconds on commodity hardware. A robot in the same kitchen, handed the same recipe and the same eggs, will fail in ways that are humbling to watch. It cracks the shell with too much force and sends yolk across the counter. It drops the pan because the gripper’s friction model underestimated the slick of cooking oil. Or it freezes for two seconds while a planner searches for a feasible trajectory, and the eggs burn.

This is not a story about robots being stupid. Modern policies trained on millions of demonstrations do remarkable things. It is a story about the difference between symbolic competence and physical competence, and Chapter 1 is about naming that difference precisely enough to engineer against it.

Three rules that do not apply in the data center

What changes when a model leaves the data center is not the size of the model. It is the set of rules the model has to satisfy. Three of them recur through every chapter of the book. Physics is non-negotiable: the moment a command leaves the policy and reaches an actuator, Newton’s laws govern what happens next, and there is no abstraction layer that hides gravity. Time is a correctness constraint, not a quality-of-service metric: an inner control loop runs at 100 to 1,000 Hz and the perception-and-policy loop on top of it at 10 to 50 Hz, and a policy that misses its budget leaves the controller with nothing fresh to act on. Errors compound through dynamics: a one-centimetre pose error at the start of a reach is a one-centimetre grasp error three seconds later and a knocked-over object eight seconds after that. Tokens are independent and replaceable. Physical errors propagate.

There is a fourth asymmetry that is less often stated. In software, producing an action is cheap relative to its consequences, and most wrong actions can be undone. In a physical system, the action itself costs energy, heat, and gearbox wear, and a great many actions are irreversible. A glass knocked off a table is not restored by an undo button. Which means that a policy that prefers to do nothing when uncertain is often the better policy — the exact inverse of the heuristic that serves us well in software.

Embodied AI is the family; Physical AI is the subset

The two terms get used interchangeably in press releases, and they are not the same thing. Embodied AI is the broad category: any agent whose decisions reference a body and an environment, including a virtual humanoid learning to walk in simulation. Physical AI is the operational subset, where the body is real, the sensors return real noise, and the actions move real mass. The distinction earns its keep because the failure modes diverge. Simulated agents reset, run a million rollouts in parallel, and can be evaluated against privileged ground truth. A real robot gets one attempt, sees only what its sensors report, and leaves a dent.

The chapter also does the unglamorous work the rest of the book depends on: separating Physical AI from classical robotics by where the competence comes from, separating it from teleoperation by where the intelligence sits, laying out the three time scales every system runs at simultaneously, and fixing the vocabulary — degrees of freedom, end-effectors, joint torques, contact-rich manipulation, 6-DOF pose, MPC, SLAM, behavior cloning, VLA — so that no later chapter has to stop and define it.

Tomorrow: none of the current design choices were arrived at in one step. Chapter 2 traces the path from Shakey and the industrial arm through behavior-based robotics and the deep-RL surge to the five phases of robotic foundation models, because each phase is best read as a response to the limits of the one before it.