Reproduce “LAFP: Preserving Latent Action Structure in Latent Policy Learning via Flow Matching” on OpenResearch

About this paper

Jiexi Lyu, Xizhou Bu, Qingqiu Huang, Chufeng Tang, Xiaoshuai Hao, Hongbo Wang, Wei Li

A framework named LAFP integrates flow matching with an inference-time interpolation mechanism to preserve the intrinsic structure of pretrained latent actions during policy learning. This method, developed by researchers primarily from Fudan University, achieved an 8.1% average increase in success rate (23.0% relative improvement) on the PROCGEN benchmark over the LAOM baseline by addressing the collapse of multimodal action distributions in latent policy distillation.

Key insights

• Preserving the intrinsic multimodal structure of the pretrained latent action space through flow matching substantially improves downstream imitation learning performance.
• The stochastic nature of flow matching policies requires an inference-time interpolation mechanism during action decoding to maintain consistency and prevent misalignment with deterministic physical actions.
• Regularizing the latent action space during pre-training (e.g., with an action decoder constraint) is critical for stable flow matching policy learning and enhanced downstream performance.

Summary by alphaXiv.