arXiv 2509.01819
ManiFlow: A General Robot Manipulation Policy via Consistency Flow Training
By Ge Yan, Jiyue Zhu, et al.
Published 2025-09-01
Discussion
Read the public discussion and references gathered around this paper.
This paper introduces ManiFlow, a visuomotor imitation learning policy for general robot manipulation that generates precise, high-dimensional actions conditioned on diverse visual, language and proprioceptive inputs. We leverage flow matching with consistency training to enable high-quality dexterous action generation in just 1-2 inference steps. To handle diverse input modalities efficiently, we propose DiT-X, a d…