This is a hybrid event with in-person attendance in Levine 307 and virtual attendance on Zoom. 

ABSTRACT

Language models became broadly capable by making reinforcement learning (RL) work via diverse pre-training, strong supervision signals, and scalable post-training algorithms. Robot manipulation policies lack this capability: demonstrations require human teleoperation, feedback is costly, and real-world reinforcement learning is too sample-inefficient to adapt policies after deployment. Rather than waiting for robot datasets to reach the required scale, my work makes robot data more informative at any scale. In this talk, I will present architectures, training objectives, and representations that do this across the robot learning pipeline to make it easier for robots to adapt after deployment. I will focus specifically on data-efficient pre-training algorithms that make adaptation easier and sample efficient, real-world RL fine-tuning algorithms for pre-trained policies. I will close with ongoing work toward robots that keep learning as foundation models advance.