Robust Hybrid Diffusion: Unifying Task Space and Joint Space for Robot Manipulation

Diffusion policies have demonstrated strong capability in learning multimodal action distributions in the context of imitation learning. While human demonstrations in general consider both task accuracy and joint configuration simultaneously, existing approaches are typically restricted to either task space or joint space, overlooking the multimodality exhibited in both spaces. To take a pushing task performed by a redundant robot as an example, it is highly desirable for the robot to push an object accurately toward its target while preserving diverse and demonstration-consistent joint-motion patterns. In addition, diffusion policies usually execute action sequences in an open-loop manner, leaving them susceptible to external disturbances (e.g., the change of payload or human intervention). To address these challenges, we propose robust hybrid diffusion (RhD), a novel imitation learning framework that unifies diffusion policies in both Cartesian space and joint space, where dynamical movement primitives are leveraged to ensure the robustness of robot actions. Evaluations on three simulated tasks and a real-world pushing task show that RhD achieves high success rates (even under external perturbations) with multimodal end-effector poses and joint configurations.