Controllable Risk Scenario Generation From Human Crash Data for Autonomous Vehicle Testing

Abstract

Ensuring the safety of autonomous vehicles (AV) requires rigorous testing under both everyday driving and rare, safety-critical conditions. A key challenge lies in simulating environment agents, including background vehicles (BVs) and vulnerable road users (VRUs), that behave realistically in nominal traffic while also exhibiting risk-prone behaviors consistent with real-world accidents. However, real-world accident data involving AVs remain extremely scarce, whereas large-scale human driving data provide a richer source of safety-critical events due to the vast cumulative driving mileage worldwide. In light of this, we propose Controllable Risk Agent Generation (CRAG), a framework that learns safety-critical behavior patterns from human crash data, combined with a nominal behavior prior, to enable agents to act adversarially toward the AV while remaining physically plausible. Extensive experiments show that CRAG improves diversity compared to existing baselines, while also enabling controllable generation of risk scenarios for targeted and efficient evaluation of AV robustness.