Generative policies that learn multimodal actions in one step

Standard RL policies face a tradeoff: Gaussian policies are fast but struggle with multimodal action distributions, while generative policies handle complex behaviors but require iterative sampling. This work proposes Stochastic MeanFlow Policies (SMFP), which use a learned transformation of Gaussian noise to generate expressive, multimodal actions—retaining tractable entropy and one-step efficiency. Trained via mirror descent with entropy regularization, SMFP improves performance over both conventional and generative baselines across seven MuJoCo benchmarks while maintaining fast inference.