Human Behavior-Informed Crash Scenario Generation with Real-World Crash Priors for Autonomous Vehicle Safety Evaluation

2026-10-03

AI Summary

This study addresses the limited reliability of existing autonomous driving safety evaluation methods, which struggle to faithfully reproduce pre-crash behavioral dynamics and accident distributions. We propose CrashSim, a framework that integrates human driving behavior priors with generative AI, leveraging large language model-assisted diagnostics and multi-agent simulation to tackle high-fidelity, scalable accident scenario generation under data-scarce conditions. Based on this framework, we construct the nuCrash dataset to enable closed-loop evaluation of planning algorithms. Experimental results demonstrate that our approach accurately replicates real-world collision dynamics, effectively differentiates the safety capabilities of various planners, and provides targeted guidance for their improvement.

Abstract

Reliable safety evaluation of autonomous vehicles (AVs) is essential to improving road safety, yet it depends critically on realistic simulation of rare crashes. Existing crash scenario generation methods can increase collision occurrence, but often fail to realistically reproduce how crashes evolve before impact or the distribution of crash types observed in the real world. Here, we present CrashSim, a human behavior-informed crash scenario generation framework that uses real-world crash priors to guide generative multi-agent traffic simulation for more reliable AV safety evaluation. These priors capture how real-world crashes evolve before impact and how different crash types are distributed, allowing limited crash data to guide realistic and scalable scenario generation across naturalistic driving contexts. We evaluate CrashSim against competing methods, showing that it more closely reproduces real-world pre-impact behavior, collision dynamics, collision geometry and crash-type distributions. We further use CrashSim to construct nuCrash dataset, containing over 4,000 crash and near-crash scenarios. Closed-loop evaluation of five AV planners shows that nuCrash more effectively exposes differences in planner safety capabilities than nuScenes. An LLM-assisted evaluation agent further analyzes planner failures to provide capability-level diagnoses and targeted improvement guidance. Together, CrashSim enables realistic and scalable crash generation for more informative AV safety evaluation.

Problem

• Autonomous Vehicle Safety Evaluation
• Crash Scenario Generation
• Real-World Crash Priors
• Pre-impact Behavior
• Crash-Type Distribution

Innovation

• Crash Scenario Generation
• Real-World Crash Priors
• Multi-Agent Traffic Simulation
• Autonomous Vehicle Safety Evaluation
• LLM-Assisted Evaluation