Bridge Policy: Visuomotor Policy Learning via Stochastic Optimal Control

Zhaoyang Liu, Mokai Pan, Zhongyi Wang, Kaizhen Zhu, Haotao Lu, Jingya Wang, Ye Shi

Keywords: visuomotor policy learning, imitation learning, generative model

TL;DR: BridgePolicy learns visuomotor policies without conditioning by bridging observations to actions via diffusion bridges.

Abstract: Imitation learning has been widely used in robotic learning, where policies are derived from expert demonstrations. Recent advances leverage generative models, such as diffusion and flow-based methods, to better capture multi-modal action distributions and temporal dependencies. However, these approaches typically impose conditioning during the forward and reverse process, which inevitably introduces manifold deviation and estimation error. In this work, we propose BridgePolicy, a condition-free generative visuomotor policy that explicitly incorporates observations into the forward process through a diffusion bridge formulation grounded in stochastic optimal control. By sampling actions from observation distributions instead of random noise, BridgePolicy reduces stochasticity and achieves more controllable policy behaviors. However, directly bridging observations to actions poses new challenges, as the action distribution may exhibit mismatched data shape, and the robot observations are inherently multi-modal. In contrast, the diffusion bridge can only connect one-to-one distributions with the same shape. To address the challenges of aligning distributional endpoints and handling multi-modal robot observations, we design a semantic aligner for distribution shape alignment, and a modality fusion module for unifying robot states and visual inputs. Experiments across 52 tasks on 3 benchmarks and 4 real-world tasks demonstrate that BridgePolicy consistently outperforms state-of-the-art generative policies.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 4991