From Diffusion to Flow: Fast and Coherent Policy Distillation via Probability Flow ODEs

Jing Wang, Xihua Wang, Li cheng

Keywords: knowledge distillation, robot policy learning, diffusion generative model

TL;DR: We introduce OneFlow-DP, a score-aligned flow-matching distillation method that preserves the stochasticity of diffusion policies while achieving single-step inference with 80× faster speed and higher performance.

Abstract: Diffusion Policies (DPs) have achieved state-of-the-art performance in visuomotor control by modeling multimodal action distributions through iterative denoising. However, inference requires hundreds of sampling steps, making real-time deployment impractical. The existing distillation methods compress DPs into single-step students but typically lose trajectory-level stochasticity and temporal coherence, which is key to DPs’ generative power.

Submission Number: 12