Coupled-Layer Diffusion for Kinodynamic Trajectory Generation

Abstract

Diffusion models have been promising for trajectory generation tasks due to their ability to capture multimodal behaviors, but their closed-box nature and unconstrained denoising often undermine kinodynamic feasibility. We propose a novel Coupled-Layer Diffusion Model (CLDM) that explicitly embeds a kinodynamic model, enhancing prior knowledge integration and improving trajectory generation performance. CLDM introduces two coupling pathways, resulting in more interpretable, efficient, and well-structured information flow. An intra-layer coupling pathway is established across coupled denoising layers, explicitly embedding prior kinodynamic information to promote information transmission during the denoising phase. This modification improves the quality of generated trajectories, particularly in terms of kinodynamic feasibility. In parallel, an inter-predictor coupling pathway that connects the noise predictors helps couple state and control information, facilitates gradient propagation, and improves training efficiency. By incorporating kinodynamic priors, we further demonstrate a more deterministic initial noise distribution and a reconstruction step that better aligns with kinodynamic feasibility. Comprehensive comparative simulations and real-world validations demonstrate faster convergence and improved generation quality. In particular, the generated trajectories exhibit superior feasibility, outperforming baselines.