Flow-Guided Policies: Overcoming Diffusion Limitations for Robust Robot Imitation Learning

Chanhyuk Jung, Sangwon Kim, Kwang-Ju Kim, Dasom Ahn, Joonki Baek, Sungkeun Yoo, Byoung Chul Ko; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 2507-2512

Imitation learning empowers robotic agents to acquire complex skills from expert demonstrations. While Diffusion Policy (DP) has recently emerged as a promising approach for capturing multimodal action distributions, its application has largely been confined to relatively simple manipulation tasks. In this work, we introduce Conditional Flow (CoF) Policy, a novel method that leverages the principles of Flow Matching to address the limitations of Diffusion Policy, enabling more complex and contact-rich robotic tasks. CoF Policy explicitly models a continuous flow that transforms suboptimal actions toward expert-like behaviors, guided by a learned vector field. Our extensive experiments demonstrate that CoF Policy consistently outperforms DP in modeling diverse multimodal action distributions and executing high-quality behaviors. Crucially, CoF Policy maintains its high performance with significantly reduced sampling steps, a desirable property for real-time robotic applications. This research marks the first instance where flow matching demonstrably exceeds the performance of diffusion-based policies in challenging imitation learning scenarios. We further validate CoF Policy through applications in simulation and real-robot case studies.