DexMP: Diffusion-Guided Motion Planning for Collision-Free Arm-Hand Manipulation in Cluttered Environments

Dexterous manipulation in cluttered scenes is challenging due to the coupled constraints of arm kinematics, multi-finger hand posture, self-collision avoidance, and contact feasibility. Compared with gripper-based manipulation, arm-hand planning involves a significantly larger hybrid discrete-continuous search space, where unsuitable pre-grasp postures may invalidate otherwise feasible motions. This paper presents a two-stage arm-hand planning framework that combines a conditional diffusion model with optimization-based refinement to generate collision-free and dynamically feasible trajectories. The diffusion model is conditioned on the robot state, task goal, scene context, and discrete hand-pose priors to generate high-quality warm-start trajectories, which are subsequently refined by an Optimal Control Problem (OCP) solver under dynamics and collision constraints. We further compare Slot Attention and point-cloud encoding for scene representation, and experimental results show that point-cloud encoding better preserves clutter geometry and improves planning stability. A compact hand configuration library is additionally introduced to provide stable pre-grasp postures for dexterous execution. Experiments on cluttered shelf-like manipulation tasks demonstrate that the proposed method outperforms representative baselines in success rate, trajectory quality, and planning efficiency.