Author’s Note

When a server returns an error for a page request, the reader can try again. A robot approaching an obstacle keeps moving while its computer decides whether to brake. Restarting the computation cannot recover the distance already traveled.

In traditional computing, software operates behind glass: an unhandled exception or late result produces a log entry, a dropped frame, or a retry button. In Physical AI, software commands matter, momentum, and energy. The unit of engineering is the action crossing the causal boundary. Delay is paid in millimeters, and failure results in cracked composite, burned motor coils, or physical harm.

To build machines that survive this reality, we must give up the idea that a high-capacity network can drive a physical plant directly. In the machine this book builds, the learned model proposes; only the permission path permits, and only for as long as its evidence holds.

I write these chapters as an active systems researcher exploring how the laws of computing intersect with the laws of Newtonian mechanics, thermal dissipation, and functional safety. Concrete systems tasks, hardware measurements, and first-principles napkin math guide our progression; formal continuous-control derivations and proof machinery provide optional depth in the appendices.

This book is an open, working notebook. Step across the causal boundary with me, and let us explore how to engineer machines that sense, act, and survive in the physical world.

— Vijay Janapa Reddi
Cambridge, Massachusetts
2026