Early Career Keynotes

Negar Mehr

Title: How Can Robots Learn to Interact and Coordinate with Others?

Abstract: Robots in the real world rarely act alone. They share spaces with other robots and people, so they need to learn how to interact and coordinate. This talk explores why that is challenging and how we can help robots succeed in interactions. I’ll begin with multi-agent reinforcement learning, where challenges such as credit assignment make coordination difficult to learn, and show how foundation models can help coach robots. Coordination can also be learned from demonstrations rather than trial and error. I’ll discuss why imitating interacting agents differs from single-agent imitation, and how it can still build on single-agent methods. Finally, I’ll turn the question around: How can ideas from multi-agent learning help a single robot learn? I’ll show how continual learning can be formulated as a consensus optimization problem which offers a natural way to balance learning new tasks with retaining previously learned skills.

Bio: Negar Mehr is an Assistant Professor of Mechanical Engineering at UC Berkeley, where she directs the Berkeley Intelligent Control (ICON) Lab. Her research focuses on multi-agent interactions. Her lab develops learning-based methods grounded in game theory and control to help robots interact safely and intelligently with people and other autonomous agents. She received her Ph.D. in Mechanical Engineering from UC Berkeley in 2019 and was a postdoctoral scholar in Aeronautics and Astronautics at Stanford University. Before returning to Berkeley, she was an Assistant Professor of Aerospace Engineering at the University of Illinois Urbana-Champaign. She has received the ONR Young Investigator Award, the AFOSR Young Investigator Award, the NSF CAREER Award, the ASME Rising Star Award, and the IEEE ITSS Best Ph.D. Dissertation Award.

Roberto Martin-Martin

Title: Learning Without Starting Over: Expanding What Robots Can Learn in the Real World

Abstract: Two robots can be equally capable today yet very different to teach tomorrow. As foundation models expand what robots can already do, an equally important question is how to expand their learning potential: the capabilities they can acquire next with the limited experience and supervision available to them. In this talk, I will argue that expanding this potential requires asking not only how robots learn, but what they should learn and reuse to acquire new capabilities safely and efficiently through real-world interaction. I will present selected learning systems for mobile manipulation that address this question in three ways: learning reusable action spaces that support safer and more efficient physical exploration; using human demonstrations and corrections to guide real-world refinement; and selecting or collecting experience according to its contribution to policy improvement. I will close with a perspective on memory and continual learning as steps toward making these improvements persist across tasks and deployments. Together, these directions aim to expand the learning frontier by preserving useful knowledge while enabling the changes that new tasks require.

Bio: Roberto Martín-Martín is an Assistant Professor of Computer Science at the University of Texas at Austin, where he directs the Robot Interactive Intelligence Lab (RobIn). He builds robotic systems that learn in the real world, studying how robots can acquire useful skills from people and physical interaction, learn safely from limited experience, and keep improving continuously over time. He pursues these questions through mobile and whole-body manipulation, contact-rich interaction, and long-horizon tasks in everyday environments. Alongside his work on physical robots, he has helped create simulation platforms and benchmarks that let researchers develop and test robot learning at scale, including Robosuite, iGibson, BEHAVIOR, and OopsieVerse.

He earned his Ph.D. at TU Berlin under Oliver Brock and conducted postdoctoral research at the Stanford Vision and Learning Lab with Fei-Fei Li and Silvio Savarese. His work has received the RSS Best Systems Paper Award, ICRA Best Paper Award, and IROS Best Mechanism Award, and he was part of the winning team of the first Amazon Picking Challenge. His individual honors include the NSF CAREER Award, Amazon Faculty Award, Meta Research Award and TRI Young Faculty Researcher Award, along with recognition as an RSS Pioneer and early-career distinctions from AAAI and IJCAI. He also serves the robotics community as Chair of the IEEE Technical Committee on Mobile Manipulation and as a co-founder of QueerInRobotics.

Hang Zhao

Title: Organizing Physical AI Across Timescales

Abstract: The physical world does not pause while a robot computes. Frontier models can now drive robots, yet they take seconds to choose a motion while contact unfolds in milliseconds. Rather than dividing a robot’s intelligence by function, I argue for dividing it by timescale, from S2 reasoning over seconds, through S1 policies at tens of hertz, to S0 control at hundreds of hertz, with each layer still learned end to end. This raises two problems. The first is speed within each layer. I will show three ways to keep S1 policies fast: learning world models from video during training while skipping video generation at execution; training with chain-of-thought supervision while acting without it; and streaming action tokenizers that execute as they are generated, keeping the policy in step with a real clock. The second is coordination across layers, through interfaces that carry intent down and feedback up. Finally, I will discuss open questions, including what S2 can express to S1; what S1 must pass to S0 beyond motion; and whether a learned policy can know when it is unsure.

Bio: Hang Zhao is an Assistant Professor at IIIS, Tsinghua University, and a co-founder of Galaxea AI, a robotics startup. His research spans foundation models for robot manipulation, agile whole-body locomotion, and vision-centric autonomous driving. Before joining Tsinghua, he was a research scientist at Waymo. He received his Ph.D. from MIT, advised by Professor Antonio Torralba.