One Diffusion Model, Two Roles: Guided Trajectory Planning and Safety-Critical Scenario Generation in Closed-Loop Simulation

To improve the closed-loop performance of nuPlan autonomous driving development, researchers proposed a single-pretrained diffusion traffic model that can simultaneously perform both ego motion planning and generation of critical safety scenarios within the closed loop. The study introduced a single-stream and dual-stream (SSDS) diffusion-Transformer decoder to fuse scene context through a joint attention mechanism; it also proposed a training-free DAPSE guidance scheme by injecting arbitrary energy functions at the clean sample level to avoid first-order approximation errors and eliminate the auxiliary network. In the nuPlan closed-loop simulation with an independent black-box planner, this model was used as a controllable scenario generator, utilizing guidance during inference to direct selected agents towards critical safety behaviors such as aggressive切入 and front-vehicle braking for closed-loop evaluation. Results showed that the generated scenarios revealed hidden failure modes under standard benchmarks; although the planner based on SSDS was stronger in nominal performance, it exhibited greater performance degradation in these challenging scenarios, proving that benchmark superiority does not necessarily translate into robustness.