Скачать или смотреть One Step Is Enough: Dispersive MeanFlow Policy Optimization (DMPO)

DMPO: Breaking the Speed-Performance Trade-off in Robot Policies | One-Step 120Hz+

The Problem: Existing generative policies face a dilemma—
• Multi-step methods (DPPO, ReinFlow): strong performance, but too slow
• One-step distillation (CP, 1-DP): fast, but unstable due to representation collapse
Fast OR good — must we choose?

Our Solution: DMPO breaks this trade-off via architecture-algorithm co-design—
• Lightweight architecture: 1.78M params (48× smaller than ViT-Base)
• MeanFlow: mathematically rigorous one-step generation, no distillation needed
• Dispersive regularization: prevents collapse, ensures stable one-step inference
• RL fine-tuning: breaks the imitation learning ceiling, surpasses expert data

These form a positive loop: dispersive reg → stable one-step → efficient RL → better policy → fast AND good.

Results: 5-20× speedup, 120Hz+ real-time control, SOTA on RoboMimic/Gym, Franka robot deployment.

Project: https://guowei-zou.github.io/dmpo-page/