Yifan Kang

I’m a junior at MIT, double-majoring in Artificial Intelligence (Course 6-4) and Mathematics (Course 18). I’m advised by Leslie Kaelbling in the Learning & Intelligent Systems Group.

Robots deployed in the real world encounter novel environments and unseen tasks beyond their training distribution, often leading to policy failures. I want to build robots that can turn these experiences into new capabilities—not just execute what they learned during training. My research asks how robots can adapt after deployment: infer useful structure from a few demonstrations, diagnose failures through interaction, and revise their behavior without collecting large datasets or retraining from scratch.

Currently, I’m working on robot-use agents that improve their policies from demonstrations. In ENCORE, I explored how a coding agent can use demonstrations directly in context to generate executable manipulation programs and refine them through execution feedback. In KeyFlow, I explored how an agent can optimize a non-parametric flow field constructed from a few demonstrations by revising the keypoints and retrieval programs that define it.

Outside the lab, I enjoy long-distance running, watching anime, rhythm games (Arcaea and Maimai), and singing.