Diffusion-Enhanced Tree Planning for Autonomous Driving

Abstract

In highly interactive urban driving, decision making is often naturally multi-stage, and decisions at different stages can lead to different reactions from surrounding vehicles. This calls for stage-wise evaluation and selection. Tree-based planning naturally supports multi-stage search and evaluation, but fixed stage partitions can introduce discontinuities at stage boundaries, leading to piecewise and unnatural trajectories. To address these issues, we propose a diffusion-enhanced tree planning framework that refines a trajectory tree in continuous space and enables single-step inference by directly predicting the denoised trajectory. The diffusion model refines ego candidates and outputs candidate-conditioned agent predictions and human-likeness scores for ranking. We then model and evaluate interaction risk directly on the tree by seeding collisions and propagating risk bidirectionally with temporal decay, together with spatial risk sharing across clustered nodes. Finally, the optimal trajectory is selected by jointly considering risk, human-likeness, and progress. Experiments show that our method produces smoother and more natural trajectories. It also achieves a better balance between safety and efficiency in closed-loop evaluation.