3 Questions: Gioele Zardini on why robots need both classical engineering and AI
MIT CEE ·
A new textbook co-authored by MIT assistant professor Gioele Zardini brings classical robotics and modern AI together and makes autonomous robotics accessible to learners everywhere. Researchers at Stanford University and MIT have published a new textbook, Principles of Robot Autonomy, that opens autonomous robotics learning to everyone. The book brings together classical robotics and modern […]
Researchers at Stanford University and MIT have published a new textbook, Principles of Robot Autonomy, that opens autonomous robotics learning to everyone. The book brings together classical robotics and modern AI, two increasingly intertwined approaches, in a textbook available in print and freely online . The book was inspired by the course of the same name that Marco Pavone P hD ‘10, professor of aeronautics and astronautics, has taught at Stanford University since 2017. Pavone, also an MIT alumnus, co-authored the book with four former members of his Stanford lab including Gioele Zardini, the Rudge (1948) and Nancy Allen Career Development Professor and assistant professor of civil and environmental engineering and principial investigator in the Laboratory for Information and Decision Systems (LIDS).
At MIT, Zardini is bringing this material into his teaching of course 1.104 Sensing and Intelligent Systems , and in part, 1.144 Applied Category Theory for Engineering Design . Students learn to reason compositionally about the design of complex systems and to see how the choices made for one component shape the rest.
In this interview, Zardini discusses why physical AI is gaining the spotlight, why now is a critical time for the field, and why engineers need to learn both classical robotics and AI.
Q1: Physical AI is getting a lot of attention right now. What do you mean by robot autonomy and physical AI, and why is this a critical moment for the field?
A: By robot autonomy, I mean the ability of a robot to perceive its environment, reason about what to do, and act on that decision without a human in the loop at every step. In the book, we organize the material around a simple cycle we call See–Think–Act. Physical AI is the broader term the community now uses for artificial intelligence that operates in the physical world. The distinction matters. A language model that makes a mistake produces a wrong sentence. A robot that makes a mistake can damage property or hurt someone.
I believe this is a critical moment for at least three reasons. First, autonomy has left the lab. Robotaxis carry passengers in several cities, fleets of robots move goods in warehouses, and drones inspect infrastructure. Second, advances in machine learning, and in particular foundation models, are giving robots capabilities that seemed out of reach only a few years ago, such as understanding scenes they have never seen or following instructions given in natural language. Finally, because these systems now operate at scale and around people, the central question is shifting. It used to be whether a robot can perform a task. Now it is whether it performs it reliably, safely, and efficiently, every time. Answering that question requires engineers who understand the whole system, and this is what we set out to teach.
Q2: AI models are increasingly learning from data, such as self-driving cars trained on human driving. Yet Principles of Robot Autonomy argues that classical robotics still matters. How do these two approaches work together, and can you give an example of where a purely data-driven system might fall short without classical methods?
A : The two approaches are complementary, and the book is built on that premise. Learning is most powerful where the world is hard to describe with equations, such as recognizing a pedestrian in a camera image or anticipating how a cyclist will behave at an intersection. Classical robotics is most powerful where we understand the structure of the problem. We know the physics of a vehicle, we know how to fuse noisy sensor measurements into a consistent estimate of where the robot is, and we know how to compute motions that respect physical limits and avoid collisions. Modern autonomy stacks combine both. A self-driving car, for instance, relies on learned perception and behavior prediction, and on model-based state estimation, planning, and control.
Consider a car trained only to imitate human driving. Such a system learns what typical drivers do in typical situations, and it can perform remarkably well on most roads. However, a purely data-driven approach faces at least three difficulties. First, the situations that matter most for safety are rare, so they are poorly represented in the data. Think of a traffic light that is out, a car pulling out from behind a parked truck, or a police officer directing traffic by hand. Second, errors compound. Each slightly imperfect action takes the car a little further from the situations it saw during training, where its predictions become less reliable. Finally, a learned policy offers no built-in guarantee that its decisions respect the physics of the vehicle or the constraints of its environment, and it is often hard to explain why it chose a particular action.
Classical methods address each of these gaps. Physical models of the vehicle keep planned motions feasible, regardless of what the data contains. Principled state estimation tells the system how uncertain it is, so that it can act more conservatively when it should. Safety layers grounded in control theory can check, and if needed correct, a learned decision before it is executed. More broadly, classical methods provide the baselines, the metrics, and the theory that let us evaluate whether a learned system is doing the right thing. Taken together, the most capable systems use learning where it brings flexibility, and classical methods where they bring structure and guarantees.
Q3: You’re now teaching and doing research in this area at MIT. Why is it important to train the next generation of engineers in both classical robotics and AI, and how does that connect to your work in CEE and LIDS? What do you hope students take away from the book, whether they’re at MIT or learning on their own online?
A: The engineers we train today will design, build, and certify the autonomous systems of the coming decades. If they know only classical methods, they will miss some of the most powerful tools of recent years. If they know only learning, they will struggle to understand why a system fails on real hardware, or how to guarantee that it will not. The goal is to give students the judgment to choose the right tool for each part of a problem, and the ability to combine these tools into a system that works.
This systems perspective is at the core of my research at MIT, at the intersection of LIDS and CEE. Autonomy is rarely deployed as a single robot. It appears as fleets of robotaxis serving a city, teams of robots operating in a warehouse, or autonomous vessels moving through a port. In these settings, the capabilities of each agent, the way agents coordinate, and the infrastructure they rely on are tightly coupled. Even the best algorithm for a single robot is of limited value if the fleet has the wrong size, if the robots are poorly matched to their tasks, or if the infrastructure was designed without them in mind. In my group, we develop the theory and the computational tools to co-design such multi-agent systems as a whole, deciding jointly on the hardware, the autonomy software, the coordination strategy, and the surrounding infrastructure, rather than optimizing each piece in isolation. LIDS brings the foundations in control, optimization, and decision-making. CEE brings the large-scale systems, such as transportation and logistics networks, where these design decisions have real impact. The principles in the book are the building blocks of this systems-level view.
I hope readers take away three things. First, a principled understanding of how the components of an autonomy stack fit together, and why they are designed the way they are. Second, hands-on experience. Chapters come with Python notebooks that run directly in the browser, so readers can implement the algorithms, tune them, and break them. Finally, a sense that this field is open to them. The full book and the exercises are freely available online, we are progressively releasing the Stanford lecture recordings, and each chapter includes Ask the Book, an AI study assistant that answers questions grounded in the chapter being read. Whether you are an MIT student, an engineer moving into robotics, or simply curious about how robots work, you can start learning today, at no cost.