WarmPrior: Straightening Flow-Matching Policies with Temporal Priors

2026-05-13

AI Summary

This work addresses the issue of curved probability paths in flow-matching-based generative robot control, which arises from the use of a standard Gaussian source distribution and often degrades task success rates. To mitigate this, the authors propose WarmPrior—a time-aware prior that leverages recent action history to replace the conventional Gaussian source distribution, yielding straighter probability paths and more efficient exploration. WarmPrior represents the first approach to explicitly incorporate temporal priors into source distribution design, significantly enhancing both sample efficiency and final performance in behavioral cloning and reinforcement learning. Experimental results demonstrate consistent improvements in task success rates and policy stability across multiple robotic manipulation benchmarks.

Abstract

Generative policies based on diffusion and flow matching have become a dominant paradigm for visuomotor robotic control. We show that replacing the standard Gaussian source distribution with WarmPrior, a simple temporally grounded prior constructed from readily available recent action history, consistently improves success rates on robotic manipulation tasks. We trace this gain to markedly straighter probability paths, echoing the effect of optimal-transport couplings in Rectified Flow. Beyond standard behavior cloning, WarmPrior also reshapes the exploration distribution in prior-space reinforcement learning, improving both sample efficiency and final performance. Collectively, these results identify the source distribution as an important and underexplored design axis in generative robot control.

Problem

• generative policies
• flow matching
• source distribution
• robotic manipulation
• temporal priors

Innovation

• WarmPrior
• flow matching
• temporal prior
• generative robot control
• rectified flow