Search papers, labs, and topics across Lattice.
This paper introduces Mean Flow Policy Optimization (MFPO), a reinforcement learning algorithm that uses MeanFlow models as a more efficient alternative to diffusion models for policy representation. MFPO addresses the challenges of action likelihood evaluation and soft policy improvement specific to MeanFlow policies within a maximum entropy RL framework. Experiments show that MFPO achieves comparable or better performance than diffusion-based RL methods on MuJoCo and DeepMind Control Suite benchmarks, while significantly reducing training and inference time.
Ditch diffusion models: MeanFlow policies offer a faster, leaner path to high-performing reinforcement learning agents.
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.