Diffusion-Guided Trajectory Priors for Efficient Robotic Manipulator Motion Planning in Dynamic Environments

Abstract

Real-time trajectory planning for 6-DoF robotic manipulators in environments containing moving obstacles remains an open challenge because planners must simultaneously satisfy collision avoidance, joint limits, and strict latency constraints. Existing diffusion-based planners suffer from three structural limitations: single-resolution denoising that forces global topology and local geometry to compete for model capacity within a single pass; static obstacle snapshots applied uniformly across all trajectory waypoints, which is incorrect because the obstacle moves by the time the arm reaches waypoints several seconds into the future; and scalar collision guidance that cannot localize corrective effort along the trajectory. In this paper, a hierarchical conditional diffusion framework is proposed for structured joint-space trajectory generation, which is conditioned on start and goal positions, static workspace geometry, and dynamic obstacle motions. To guarantee temporal consistency, a spatio-temporal cross-attention module is employed to align each waypoint query with predicted obstacle state at its corresponding execution time. Furthermore, a learned GuidanceScheduler is proposed that uses a waypoint-wise collision penalty profile, enabling temporally directed guidance during denoising instead of a scalar penalty. The framework is trained on 10,800 expert trajectories for the Kinova J2S6S300 manipulator and evaluated on 1200 held-out scenarios. It achieves 93.5% planning success versus 82.5% for the best learning-based baseline, Hierarchical Diffusion Policy (HDP), an absolute improvement of 11 percentage points, and 95.0% when combined with ST-CHOMP post-processing, at 243 ms inference latency.