A new reinforcement-learning method lets a control policy pick its next move in a single step, without giving up the ability to hedge between multiple good options.
Researchers describe Stochastic MeanFlow Policies (SMFP), which pair a MeanFlow-based mapping from noise to actions with a Gaussian reparameterization trick. That combination lets the policy capture multiple distinct ways of solving a task - what the researchers call multimodal action modeling - while still producing an analytic lower bound on entropy, a measure of how random the policy's choices are. The team used that entropy bound to drive exploration directly, combining it with value guidance and advantage-weighted regression in an entropy-regularized mirror-descent training setup. Tested across seven MuJoCo physics-simulation benchmarks, SMFP matched or beat both Gaussian and other generative baselines, despite needing only one step to sample an action.
That one-step part is the real story. Diffusion-style generative policies are good at representing varied, multimodal behavior, but they usually need many iterative denoising steps to produce a single action - a nonstarter for real-time control. Gaussian policies sample instantly but collapse multiple good strategies into one blurry average. SMFP's pitch is getting both at once: fast sampling with an explicit, tunable knob for exploration, something most generative-policy methods lack.
It's still simulation-only, run on standard MuJoCo tasks rather than physical robots, so treat "effective exploration" as a benchmark result, not a guarantee it survives contact with messier real-world sensors and dynamics.