In robotic manipulation, the same observation often admits multiple feasible actions — a robot may grasp a cup from the side or from above, for instance. Generative policies such as diffusion and flow matching can capture this multimodality naturally, but the paper notes that which mode is executed is left to random chance in standard formulations.
MOSS (Mode-Oriented Source Shaping) targets exactly that gap. The method shapes the source distribution used in flow-matching policies so that the model can be oriented toward a particular action mode, rather than leaving the choice to the stochastic sampling process. The title suggests the intervention happens at the source level of the flow, before the policy refines it into an action.
The source abstract is truncated, so details on the exact shaping mechanism and experimental benchmarks are not available in the provided text. The paper is listed as arXiv:2610.22193v1.