Abstract
Safety-critical scenarios in natural driving data exhibit a distinct long-tail characteristic and scarcity, which limits the construction and analysis of safety-critical traffic scenarios. This paper addresses the problem of generating multi-vehicle dangerous trajectories on highways using naturalistic trajectories from the highD drone dataset, and proposes a multi-agent diffusion policy (MADP) fine-tuning framework. Firstly, a conditional diffusion model is used to learn the joint acceleration distribution, and the historical motion, map lanes, and vehicle interaction information are integrated through unified spatiotemporal encoding and hierarchical conditional fusion. During reward-driven fine-tuning, a clipped update inspired by proximal policy optimization uses a dimension-normalized denoising-chain score computed from stochastic denoising transitions. A role-weighted noise reconstruction anchor limits shifts in the learned behavior distribution. The scene reward considers the dangerous approach of the ego vehicle and the attacking vehicle, the trajectory characteristics of the restricted movement of the ego vehicle, as well as the collision side effects of kinematic violations and non-ego vehicle pairs. Finally, the experimental results on the highD dataset show that the proposed method can generate multi-vehicle trajectories including target collisions and low safety gaps; the ablation results reflect the influence of dangerous rewards and supervised anchors on risk and trajectory quality. At the same time, risk enhancement comes with certain deviation of behavior distribution and smoothness cost. Based on this, this paper discusses the risk and quality trade-off in the generation of dangerous trajectories.