Most modern robot policies are frozen at training time. A diffusion policy trained on a few hundred demonstrations performs well on tasks it has seen, but cannot meaningfully grow from its own experience or absorb new skills without a costly offline retraining loop. This is a fundamental obstacle to building robots that get better the longer they are deployed. The broad goal of this SuperUROP is to explore methods that let a robot continually learn – accumulating skills, refining old ones, and adapting to novel situations from a small amount of supervision.

Robotics has been progressing incredibly fast in recent years. I want to deepen my research experience by working on a longer-term project in this space, building on my previous experience in Professor Kaelbling’s group. I am especially excited to get more hands-on time with real robot and to develop the skills to run a full research project.