A Note on AI Assistance

Every chapter of this book was researched, written, and rewritten by me. I also used AI tools throughout. What follows is more than the disclosure MIT Press requires of authors who use AI. It is an honest statement of how I used these tools, a model for how students and practitioners can engage with them responsibly, and a claim about what authorship means when machine learning acts as both force multiplier and sounding board.

In my previous volumes, systems and computing infrastructure was my native intellectual home. But Physical AI pushed me across the causal boundary into unfamiliar terrain: classical kinematics, nonlinear control barrier functions, motor winding thermodynamics, and high-frequency real-time bus contention. The most difficult challenge in writing this book was not drafting prose; it was conducting an exhaustive cross-disciplinary literature survey and synthesizing fields that rarely speak to one another. I used large language models extensively as interactive research partners to survey disparate literature, explore unfamiliar dynamics, brainstorm structural framings, and distill foundational first principles. These models excel at first-principles distillation—which is precisely how I naturally think about computer systems.

Yet distillation is only an initial proposal. Because these concepts live in the physical world where errors carry kinetic consequences, I relied heavily on human experts—both academic researchers in robotics and control theory, and practicing engineers at autonomous systems companies—to critique, validate, and challenge the architectures developed here. This book does not claim to be a final, closed work; it is an active, open iteration that evolves alongside a rapidly developing discipline.

The standard remains uncompromising traceability. Every derived metric, timing budget, and energy calculation in these chapters is computed directly by executable Python cells whose source code ships with the book. Quantitative system simulations are backed by the MLSysIM infrastructure modeling engine. A continuous pre-commit suite verifies dimensional units, citation integrity, and reference anchors on every change. AI tools made navigating a steep cross-disciplinary learning curve mechanically tractable; our verification infrastructure makes that work reproducible and auditable.

Authorship is thinking. Deciding which concepts belong in this curriculum and which do not; sequencing the chapters so that each abstraction earns its keep; determining where students will get stuck because I have watched them get stuck in my own classroom—these are judgment calls that no tool can make. The ideas and architectural synthesis are mine; the initial prose was often drafted, expanded, or restructured by AI before being aggressively rewritten and verified.

Machine learning systems cannot be certified the way traditional deterministic logic can. That verification gap applies to AI-assisted authoring with equal force. The standard I held this book to is the standard we must demand of every embodied machine we deploy: the tool can propose, and the tool can tutor, but only a human engineer can ratify and take responsibility.