About the Author

Vijay Janapa Reddi is the Gordon McKay Professor of Electrical Engineering and Computer Engineering at Harvard University and a Visiting Professor at ETH Zurich in the Integrated Systems Laboratory (IIS), D-ITET. He is a founding member and Vice President of MLCommons and the author of the open-source reference text Machine Learning Systems.

His research sits at the intersection of computer architecture, machine learning systems, and edge robotics, focusing on how computing hardware and learned software co-design to interact with the physical world. Over the past decade, his work has concentrated on democratizing applied systems through open-source education and reproducible benchmarks, connecting first-principles silicon constraints directly to accessible, bench-verified learning.

Why I Am Writing This Book: The First-Principles Systems Stance

Every textbook reflects the intellectual lineage, biases, and lens of its author. I want to make the viewpoint behind this book explicit.

I am not a classical control theorist, nor a mechanical kinematics specialist, and I do not pretend to be one. I bring the perspective of a computer systems architect and machine learning systems (MLSys) researcher going after first principles.

My intuition for this domain dates back to my undergraduate mechatronics capstone. As an electrical and computer engineering student, that capstone was the first time I had to integrate everything—analog circuits, sensor conditioning, microcontrollers, motor drivers, mechanical linkages, and embedded software—into a single working machine. It was the moment I realized that software cannot be designed in isolation from physical reality: motor coils heat up, power rails droop under sudden acceleration, and sensor measurements age the instant physical transduction occurs.

For over twenty years—from mobile System-on-Chip performance architecture and edge AI silicon at Google, to co-founding MLCommons/MLPerf to establish empirical benchmarking standards for AI accelerators, to pioneering the TinyML movement to compress neural networks onto milliwatt microcontrollers—my work has lived at the physical boundary where software bits meet physical silicon and energy constraints.

Physical AI brings me full circle back to that mechatronics capstone, but with a profound twist. In classical mechatronics, we closed the loop using hand-crafted, deterministic logic: PID controllers, state machines, and analytical models. In Physical AI, we are embedding high-capacity, stochastic neural networks—foundation models, diffusion policies, and learned decoders—into that exact same unforgiving loop. The engineering challenge is no longer just making the machine move; it is figuring out how to safely close the loop around an uncalibrated, stochastic brain using deterministic systems and safety principles.

That background shapes what I want students to understand. A model’s proposed action must pass through a computing system before the physical machine can carry it out. I want readers to trace that path and ask what could prevent an action from being carried out as intended. The timing of computation, the body’s capabilities, and the assumptions behind command checks all matter to that assessment. The explanations should make these interactions understandable before specialized mathematics or implementation detail is needed.

This textbook is my attempt to share that foundational systems perspective with the next generation of engineers, roboticists, and computer scientists. It is the book I wish my own students had when bridging machine learning systems to the physical world.

Intellectual Lineage and Systems Trajectory

This first-principles systems trajectory spans three major eras:

• Silicon Reality & Mobile Architecture: Grounded in computer architecture, low-power microarchitectures, and mobile System-on-Chip (SoC) design (including edge AI silicon performance architecture at Google). This work addressed the physical realities of computing: memory bus contention under high-bandwidth camera streams, thermal throttling under strict dissipated wattage envelopes, and interrupt latencies under hard real-time deadlines.
• ML Systems & The TinyML Movement: Bridging algorithmic machine learning abstractions to silicon execution. He co-founded MLPerf to establish the global standard for benchmarking ML hardware and software, and pioneered the TinyML movement, developing the Harvard CS249r curriculum and the HarvardX TinyML program (reaching over 100,000 learners globally) to bring deep neural models onto milliwatt microcontrollers.
• Closing the Sense-and-Act Loop in Physical AI: Extending this systems perspective to machines whose learned models propose physical actions. The focus is on how sensing, computation, and the body’s response interact, and on the conditions under which a proposed action can be executed.

The Kit and Laboratory Development

Development of the hands-on bench laboratories and the reference hardware platform (the Physical AI Kit on Arduino UNO Q Dual-Brain) is a collaboration with Dr. Andrea Mattia Garavagno, postdoctoral researcher in the Integrated Systems Laboratory (IIS), D-ITET at ETH Zurich.

This work aims to translate the systems ideas in the book into firmware and software exercises that students and practitioners can reproduce and examine on the bench.