FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation

• DOI: 10.1609/aaai.v39i14.33617
• Publication Date: 2025-04-11T12:24:14Z
• Authors: Qinglun Zhang; Zhen Liu; Haoqiang Fan; Guanghui Liu; Bing Zeng; Shuaicheng Liu

Robots can acquire complex manipulation skills by learning policies from expert demonstrations, which is often known as vision-based imitation learning. Generating based on diffusion and flow matching models has been shown to be effective, particularly in robotic tasks. However, recursion-based approaches are inference inefficient working noise distributions policy distributions, posing a challenging trade-off between efficiency quality. This motivates us propose FlowPolicy, novel framework for fast generation consistency 3D vision. Our approach refines the dynamics normalizing self-consistency of velocity field, enabling model derive task execution single step. Specifically, FlowPolicy conditions observed point cloud, where directly defines straight-line flows different time states same action space, while simultaneously constraining their values, that is, we approximate trajectories robot actions field within thus improving efficiency. We validate effectiveness Adroit Metaworld, demonstrating 7× increase speed maintaining competitive average success rates compared state-of-the-art methods.