Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge with Observation-Embedded Stochastic Differential Equation

The article introduces BridgePolicy, a generative visuomotor policy that integrates observations into the stochastic dynamics of the diffusion process via a diffusion-bridge formulation. This approach allows sampling to commence from an informative prior rather than random noise, enhancing control precision and reliability. Extensive experiments across 52 simulation tasks and 5 real-world tasks show that BridgePolicy outperforms existing state-of-the-art generative policies, addressing the challenge of aligning heterogeneous robotic observations with action representations.