Keywords: Flow Matching, Robot Manipulation
TL;DR: This paper presents a flow matching policy for robot manipulation, which represents a robot visuomotor policy as a conditional process of flowing random waypoints to desired robot action trajectories.
Abstract: This paper presents a new imitation learning paradigm for robot manipulation with flow matching policy. Flow matching represents a robot visuomotor policy as a conditional process of flowing random waypoints to desired robot action trajectories, by regressing vector fields of fixed conditional probability paths. We evaluate the proposed method across two simulation benchmarks and a real-world dataset with 10 tasks across Activities of Daily Living. Our extensive evaluation highlights that learning multimodal robot actions with flow matching policy leads to consistently more stable training and faster generalization than alternative diffusion-based behavior cloning methods.
Submission Number: 30
Loading