I work on robot learning for reliable long-horizon manipulation. I treat a long task as a chain of skills and study the full lifecycle of a skill: how to collect and generate the demonstrations that teach it, how to make each skill robust and safe on its own, and how to compose skills into long tasks and evaluate them.

Representative work: collecting and generating demonstrations with AR (ARCADE) and with a real-to-sim engine that turns human hand motion into 223k dexterous demonstrations with contact-force labels (GNR); keeping each skill robust to sensor corruption across vision and touch (EGR) and safe through barrier functions learned from demonstrations (SECURE); and composing and evaluating skills, by exposing how earlier skills break later ones (BOSS), linking object-centric VLA skills that hold up to these shifts (LiLo-VLA), and autonomously resetting and scoring long-horizon rollouts on real robots (HALTER).