CrashSim, a computer framework for generating safety-critical driving situations, produced simulated crashes that more closely followed real-world crash behavior than the competing generators tested in an arXiv preprint. The scenarios matched not only the final collision, but also the changing pre-collision dynamics, response timing and braking intensity. They also came closer to real-world crash statistics across both detailed and broad crash categories.
The clearest practical difference appeared when the scenarios were used to test autonomous-driving planners, the systems that choose how a vehicle should move. In closed-loop evaluation, the planners were tested through a continuing simulated sequence rather than a single isolated decision. Five planners spanning rule-based and learning-based approaches were evaluated on both nuScenes and nuCrash, a collection of over 4,000 safety-critical scenarios that included crashes and near-crashes. The normalized performance range rose from 0.08 on nuScenes to 0.58 on nuCrash. A separate measure of how much the planners differed from one another, between-planner variance, rose from 0.001 to 0.065. The nuCrash results therefore separated the planners much more clearly.
Building scenarios around real crashes
CrashSim is built around four ingredients: scene understanding, hierarchical retrieval of real-world crash priors, guidance informed by human behavior, and diffusion-based generation of trajectories for multiple agents. Its crash-prior database uses SHRP2 NDS trajectory reconstructions, assigns interactions to 16 modes derived from NHTSA crash information, and resamples trajectories at 0.5-second intervals. The design gives the generator reference patterns for how risky interactions develop, rather than focusing only on whether a collision occurs.
The source material came from 1,000 urban nuScenes scenes, each lasting 20 seconds. The study compared CrashSim with other scenario generators and compared nuCrash with SimEngine by looking at how closely their crash-type distributions lined up with real-world crashes. CrashSim had the lowest reported discrepancy across both fine-grained and high-level categories. nuCrash had smaller TV and JSD distances, the study's measures of the gap between category distributions, than SimEngine while retaining the reported behavioral and distributional realism of real-world crashes.
What the component tests changed
Two component tests illustrated how the pipeline's parts changed its output. In the version without crash-prior retrieval, the emergency-response proportion was 12%, compared with 22% when retrieval was present and 32% in the real-world reference. The comparison describes a clear change in the mix of responses represented in the generated scenarios.
In the version without the VLM agent, which supplies scene understanding, TV distance was 0.281, versus 0.093 with the agent. JSD was 0.185 versus 0.078. Together, these comparisons correspond to the framework's separate crash-prior retrieval and scene-understanding components.
Useful analysis, with a narrow reach
An automated language-model evaluation agent combined numerical results with evidence from individual scenarios to assess planner capabilities and diagnose failures. Those assessments were reported as highly consistent with the quantitative evidence, with generally high evidence sufficiency and diagnostic depth. The authors caution, however, that attribution of the underlying mechanisms behind planner failures remains relatively coarse. The analysis can organize evidence around a planner's performance, but it should not be read as a definitive account of why a failure occurred.
The work is identified as an arXiv preprint, and its main test is a closed-loop simulation involving nuScenes, nuCrash and five planners. The results therefore describe how systems compared in this benchmark, not whether autonomous vehicles would have fewer real-world crashes. The study's central contribution is a more demanding simulated setting for comparing planners, with scenarios designed to represent the build-up to a crash as well as the collision itself.
Paper data and sources
Original title: Human Behavior-Informed Crash Scenario Generation with Real-World Crash Priors for Autonomous Vehicle Safety Evaluation
Authors: Mingxing Peng, Xusen Guo, Long Chen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-03
DOI: Not available
Original paper · Full text