Flow Policy: Generalizable Visuomotor Policy Learning via Flow Matching

摘要

An effective method to improve the performance of robot imitation learning is generating sequence models such as the diffusion model, which is simulated-based and currently lacking flexibility. In this article, we introduce a novel and general methodology called flow policy to address this challenge. Specifically, the flow policy integrates flow matching, a simulated-free sequence modeling method, into robot behavior cloning. To fully unlock the potential of flow matching for imitation learning on robots, we present a set of key technical contributions, including a noise scheduler based on higher order trigonometric functions, time-varying noise, and a UDiT1d network with self-supervised learning. To validate its efficacy, we implemented the technique across several simulation benchmarks and real-world environments, achieving superior performance with fewer parameters compared to the diffusion policy. Notably, our method resulted in a 2.41% improvement in the 2-D simulation task (36 000 tests), a 6.21% improvement in the 3-D simulation task (36 000 tests), and a remarkable 34.10% improvement in the real-world task (90 tests).