Mean Flow Policy Optimization

TL;DR

MeanFlow Policy Optimization (MFPO) introduces efficient flow-based policies for reinforcement learning, achieving comparable or better performance than diffusion models with significantly reduced training and inference times.

Contribution

The paper proposes MeanFlow models for RL policies, enabling faster training and inference while maintaining high performance, and develops a maximum entropy RL framework for these models.

Findings

MFPO matches or exceeds diffusion-based RL performance.

MFPO significantly reduces training and inference time.

Experiments on MuJoCo and DeepMind benchmarks validate effectiveness.

Abstract

Diffusion models have recently emerged as expressive policy representations for online reinforcement learning (RL). However, their iterative generative processes introduce substantial training and inference overhead. To overcome this limitation, we propose to represent policies using MeanFlow models, a class of few-step flow-based generative models, to improve training and inference efficiency over diffusion-based RL approaches. To promote exploration, we optimize MeanFlow policies under the maximum entropy RL framework via soft policy iteration, and address two key challenges specific to MeanFlow policies: action likelihood evaluation and soft policy improvement. Experiments on MuJoCo and DeepMind Control Suite benchmarks demonstrate that our method, Mean Flow Policy Optimization (MFPO), achieves performance comparable to or exceeding current diffusion-based baselines while considerably reducing training and inference time. Our code is available at https://github.com/MFPolicy/MFPO.