Mechanical and Civil Engineering Seminar

Title: "Robots That Learn To Model The World By Playing"

Abstract: Predictive models of the world are fundamental to robotics — from planning, to evaluation, to reinforcement learning. However, despite decades of investment in physics-based simulation, we do not have models that can endow robots with general manipulation capabilities. In this talk, I will discuss the potential of action-conditioned video generation models to serve as general-purpose world models for robotics. Their ability to generate photorealistic observations, to simulate complex physical interactions with non-rigid objects, and to be improved with data make them an attractive alternative to traditional physics-based models. But, how should we learn such world models? I will argue that autonomous play may hold the key. Just as self-guided play is critical for children to explore and learn the dynamics of the world, autonomous robot play provides a scalable pathway for capturing the complex, long-tailed physical interactions essential for manipulation. I will show how world models learned through play demonstrate significant improvements in accuracy compared to models trained on data from human tele-operation. I will also show how the resulting world models can be used (i) as "simulators" for evaluating the reliability and safety of robot policies, and (ii) for training policies via reinforcement learning in the world model. I will end by discussing recent work on world models that know when they don't know through rigorous uncertainty quantification.

Bio: Anirudha Majumdar is an Associate Professor at Princeton University in the Mechanical and Aerospace Engineering (MAE) department, and founding co-Director of the Princeton Robotics Initiative. He also holds a 20% research scientist position at Google DeepMind in the Robotics Safety & Alignment team. Majumdar received a Ph.D. in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology in 2016, and a B.S.E. in Mechanical Engineering and Mathematics from the University of Pennsylvania in 2011. Subsequently, he was a postdoctoral scholar at Stanford University from 2016 to 2017 at the Autonomous Systems Lab in the Aeronautics and Astronautics department.

Majumdar is a recipient of the Sloan Fellowship, ONR Young Investigator Program (YIP) award, NSF CAREER award, Google Faculty Research Award (twice), Amazon Research Award (twice), Young Faculty Researcher Award from the Toyota Research Institute, Best Student Paper Award (as advisor) at the Conference on Robot Learning (CoRL), Paper of the Year Award from the International Journal of Robotics Research (IJRR), Best Conference Paper Award at the International Conference on Robotics and Automation (ICRA), Alfred Rheinstein Faculty Award (Princeton), and the Excellence in Teaching Award (Princeton SEAS).