Kinematics-Aware Diffusion Policy with Consistent 3D Observation and Action Space for Whole-Arm Robotic Manipulation
Kangchen Lv, Mingrui Yu, Yongyi Jia, Chenyu Zhang, Xiang Li†
†Corresponding Author
Published in IEEE Robotics and Automation Letters (RA-L), 2026
This paper proposes a kinematics-aware imitation learning framework for whole-arm robotic manipulation. Traditional methods using end-effector poses are insufficient for full-body control, while joint-space approaches suffer from misalignment with the actual 3D task space. We represent both robot states and actions as a set of 3D points on the arm body, creating a consistent representation across observation, action, and task spaces that naturally aligns with 3D point cloud observations. Built upon diffusion policy, we incorporate kinematics priors into the diffusion process to guarantee kinematic feasibility and use an optimization-based whole-body inverse kinematics solver for execution. Experiments in simulation and real-world demonstrate higher success rates and stronger spatial generalizability compared to existing methods.