Explainable trajectory planning for autonomous driving via dynamic topology-reasoned diffusion model

Multimodal trajectory planning is crucial for autonomous driving in complex interactive traffic, yet maintaining interaction consistency, behavioral separability, and model interpretability during generation remains challenging. To address these challenges, this paper proposes an Explainable Trajectory Planning method via a Braid-Theory-Informed Diffusion Model (BraidDiff). BraidDiff introduces braid-theory-based topology reasoning into diffusion-based multimodal trajectory planning, where interaction orders, crossing relations, and topological complexity are explicitly encoded from dynamic traffic scenes. Specifically, the method encodes agent motion history and lane geometry into braid-inspired topological representations, from which a topology reasoning module infers interaction relations and maneuver-level topology priors. These priors are progressively injected into the denoising process, guiding trajectory generation toward structurally consistent and behaviorally distinct modes. Auxiliary supervision on braid consistency, interaction-edge prediction, and topology-related mode classification further strengthens topology-aware learning under dense interactions. In addition to planning performance, BraidDiff exposes explicit topology-related intermediate variables that support structured interpretation of candidate generation and trajectory selection. Closed-loop experiments on the nuPlan benchmark show that BraidDiff achieves strong overall performance, with clearer advantages in highly interactive and challenging scenarios. Further analyses verify that explicit topological interaction modeling improves the robustness and interpretability of diffusion-based planning. Deployment experiments on an onboard domain controller further demonstrate that the proposed method achieves favorable real-time inference performance and has promising potential for engineering deployment.