AnyManip
Learning diverse robot manipulation skills through optical flow prediction.
AnyManip explores optical flow as a transferable representation of end-effector and object motion for robot manipulation. I developed a two-stage system that first predicts task-relevant optical flow with a diffusion model, then combines RGB observations, predicted flow, and proprioception to plan actions in unfamiliar environments.
This project was conducted at the Beijing Institute for General Artificial Intelligence with Dr. Siyuan Huang and Dr. Tengyu Liu.