A Similarity-Recombination-Based Enhancement Method for Kinematic Diffusion Trajectory Planning

In response to the problems of decreased rationality, reduced obstacle-avoidance safety, and insufficient kinematic feasibility of data-driven diffusion-model planners in few-shot scenarios, we propose a kinematic diffusion trajectory planning enhancement method based on similarity recomposition. By constructing a joint metric of scene similarity and state similarity, similar trajectory primitives are selected and recomposed, thereby improving the rationality and obstacle-avoidance safety of sample trajectories. Furthermore, a diffusion model with explicit kinematic embedding is constructed so that it can serve both as a recomposed-trajectory correction module and as a trajectory planner, improving the kinematic feasibility of sample trajectories and strengthening the planner’s kinematic guidance constraints. Overall, the proposed method collaboratively enhances the applicability of the planner to few-shot scenarios from two aspects: data augmentation and constraint guidance. Experimental results show that, in few-shot scenarios, the proposed method reduces the kinematic error of planned trajectories by up to 83.3% and improves the obstacle-avoidance success rate by 4.4%. The proposed method was validated through physical experiments.