Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge with Observation-Embedded Stochastic Differential Equation
Researchers have introduced BridgePolicy, a novel visuomotor policy that integrates observations directly into the diffusion process for improved robotic control. Unlike previous methods that treated observations as mere conditions, BridgePolicy embeds them into the stochastic dynamics, allowing sampling to begin from an informative prior rather than random noise. This approach enhances precision and reliability, particularly in tasks with heterogeneous robot data, by employing a semantic aligner to unify visual and action representations. Experiments across numerous simulation and real-world tasks demonstrate BridgePolicy's superior performance compared to existing generative policies. AI
IMPACT Enhances robotic control precision and reliability by integrating observations into diffusion models.