Dr. Cathy Wu serves as the Gilbert W. Winslow Assistant Professor of Civil and Environmental Engineering at MIT, while also contributing as a member of the MIT Institute for Data, Systems, and Society. During her undergraduate years, Wu conquered MIT's most challenging robotics competition. As a graduate student, she participated in the University of California at Berkeley's pioneering course on deep reinforcement learning. Now back at MIT, she's revolutionizing robot coordination in Amazon warehouses through the Science Hub, an innovative partnership between the tech giant and the MIT Schwarzman College of Computing. Beyond her academic pursuits, Wu enjoys running, creating art, crafting lattes at home, and expanding her knowledge through educational YouTube channels like 3Blue1Brown and Practical Engineering. She recently paused her busy schedule to share insights about her groundbreaking work.
Q: What inspired your journey into robotics and autonomous vehicles?
A: While my parents hoped for a medical doctor in the family, I've never been good at following instructions! My high school physics and computer science classes sparked my passion for engineering. I wanted to make as significant an impact on people's lives as a medical doctor could, just through a different path.
At MIT, I explored applications in energy, education, and agriculture, but autonomous vehicles truly captured my imagination. The statistics are compelling: approximately 94% of serious car crashes stem from human error and could potentially be prevented through self-driving technology. Beyond safety, autonomous vehicles promise to reduce traffic congestion, conserve energy, and enhance mobility for all.
My introduction to self-driving cars came from Seth Teller during his guest lecture for the Mobile Autonomous Systems Lab (MASLAB) course, where MIT undergraduates compete to build fully functional robots from scratch. Our ball-collecting robot, Putzputz, secured first place. This victory propelled me to take more courses in machine learning, computer vision, and transportation, eventually joining Teller's research lab. I also participated in various mobility-focused hackathons, including one sponsored by Hubway (now Blue Bike).
Q: You've researched ways to improve interactions between humans and autonomous vehicles. What makes this challenge so complex?
A: The complexity arises because both systems are incredibly intricate, and our traditional modeling tools fall short. Integrating autonomous vehicles into existing transportation networks represents a monumental task. For instance, predictions about autonomous vehicles' energy impact vary wildly – from reducing consumption by 40% to potentially doubling it. We need more sophisticated tools to navigate these uncertainties. My PhD dissertation at Berkeley addressed this challenge by developing scalable optimization methods for robot control, state estimation, and system design. These approaches help decision-makers anticipate future scenarios and design better systems that accommodate both human and robotic elements.
Q: How is deep reinforcement learning, which combines deep and reinforcement learning algorithms, transforming the field of robotics?
A: I took John Schulman and Pieter Abbeel's reinforcement learning course at Berkeley in 2015, shortly after Deepmind published their groundbreaking paper in Nature. They had trained an agent using deep learning and reinforcement learning to master "Space Invaders" and other Atari games at superhuman levels, creating tremendous excitement in the field. A year later, I began applying reinforcement learning to mixed traffic systems where only some vehicles are automated. I quickly recognized that classical control techniques couldn't handle the complex nonlinear control problems I was working on.
While Deep RL has become mainstream, it hasn't fully permeated robotics, which still heavily depends on classical model-based control and planning methods. Deep learning remains crucial for processing raw sensor data like camera images and radio signals, while reinforcement learning is gradually being integrated into robotic systems. I view traffic systems as enormous multi-robot networks. I'm particularly excited about an upcoming collaboration with Utah's Department of Transportation to implement reinforcement learning for coordinating vehicles with traffic signals, which could reduce congestion and carbon emissions.
Q: You've mentioned the MIT course 6.003 (Signals and Systems) and its influence on you. What resonated with you about it?
A: The perspective it offered. The course demonstrated that seemingly chaotic problems could be analyzed using common, sometimes surprisingly simple, tools. Signals undergo transformations through various systems, but what do these abstract concepts really mean? A mechanical system might transform a signal like rotating gears at one speed into a lever turning at another speed. A digital system could convert binary digits into other binary data, text, or images. Financial systems transform news through millions of trading decisions into stock prices. People constantly process signals through advertisements, job opportunities, conversations, and more, translating them into actions that influence society and others. This unassuming course on signals and systems connected mechanical, digital, and social systems, revealing how fundamental tools can cut through complexity.
Q: In your collaboration with Amazon, you're developing systems for warehouse robots to collect, sort, and deliver products. What technical obstacles are you addressing?
A: This project focuses on two main challenges: assigning robots to specific tasks and determining optimal routes to reach them. [Professor] Cynthia Barnhart's team concentrates on task assignment, while my team focuses on path planning. Both problems fall under combinatorial optimization because the solution involves selecting the best combination from numerous possibilities. As the number of tasks and robots increases, the potential solutions grow exponentially—a phenomenon known as the curse of dimensionality. Both problems are classified as NP Hard, meaning no efficient algorithm may exist to solve them perfectly. Our objective is to develop practical shortcuts that provide good solutions.
Routing a single robot for a single task isn't particularly challenging—similar to using Google Maps to find the quickest route home. This can be efficiently solved with algorithms like Dijkstra's. However, warehouses function like small cities with hundreds of robots operating simultaneously. When traffic jams occur, customers experience delays in receiving their packages. Our goal is to create algorithms that determine the most efficient paths for all robots operating in the warehouse environment.
Q: What other applications might this research have?
A: Absolutely. The algorithms we're testing in Amazon warehouses could eventually help alleviate congestion in actual cities. Other potential applications include managing aircraft on runways, coordinating drone swarms in airspace, and even controlling characters in video games. These algorithms might also be applied to other robotic planning challenges involving scheduling and routing.
Q: AI is advancing rapidly. In which areas do you anticipate major breakthroughs?
A: I hope to see deep learning and deep RL applied to solve societal challenges in mobility, infrastructure, social media, healthcare, and education. While Deep RL has gained a foothold in robotics and industrial applications like chip design, we must exercise caution when implementing it in systems that involve human interaction. Ultimately, our goal is to design systems that serve people's needs. Currently, we lack the appropriate tools to achieve this effectively.
Q: What concerns you most about AI taking on increasingly specialized tasks?
A: AI offers tremendous potential for positive impact, but it could also accelerate the growing divide between privileged and disadvantaged populations. Our political and regulatory systems could help integrate AI into society while minimizing job displacement and economic inequality. However, I worry that these systems aren't adequately prepared to handle the rapid pace of AI development and implementation.
Q: What's the most recent book that left a significant impression on you?
A: "How to Avoid a Climate Disaster," by Bill Gates. I was deeply impressed by Gates' ability to take an overwhelmingly complex topic and distill it into concepts accessible to everyone. His optimistic approach inspires me to continue exploring how AI and robotics can contribute to preventing climate catastrophe.
