Autonomous Vehicle Safety Performance in Mixed Traffic: Insights from NHTSA Crash Data

The safe deployment of autonomous vehicles (AVs) depends on the ability of automated driving systems (ADS) to handle rare, complex, and safety-critical “edge cases.” This study develops a novel framework for identifying and analyzing such scenarios using the National Highway Traffic Safety Administration (NHTSA) ADS crash dataset. Two complementary approaches are applied. First, large language models (LLMs) are used to directly analyze crash narratives, identifying edge cases through high-risk keyword and phrase detection and anomaly-based rarity analysis. Second, LLMs are employed to extract structured variables from narrative fields, which are then analyzed using hierarchical clustering to systematically isolate unusual crash groups. Edge cases are characterized by a higher prevalence of unusual crash partner behaviors, non-motorist involvement, roadway anomalies, and disengagement of AV systems, highlighting their distinct and atypical nature. The findings underscore the importance of focusing AV evaluation on rare, high-risk scenarios that challenge ADS performance. The study advances AV safety research and can provide a foundation for refining testing protocols, safety standards, and regulatory frameworks to better capture the operational limits of AVs.