Revealing the behavioral architecture of autonomous vehicle crashes: an interaction-based typology using real-world...

Over recent years, hundreds of crashes have been recorded during on-road testing of autonomous vehicles (AVs), underscoring the importance of understanding how these systems behave alongside human drivers in mixed traffic. Most previous research has focused on geometric crash outcomes, offering limited characterization of the paired vehicle motion states preceding impact. This study examines verified California DMV OL-316 crash reports (2024-2025) to construct an Interaction-Based Crash Typology (IBCT) that systematically describes AV-human motion configurations based on paired discrete pre-crash motion states recorded in static crash reports, rather than continuous, real-time vehicle-to-vehicle interactions. Through descriptive statistics and Apriori association-rule mining, twenty-four unique interaction patterns were identified, with the configuration AV stopped × OV proceeding straight representing the largest share (27.6%). The first hypothesis (H1) was supported descriptively, showing recurrent co-occurrence patterns between AV and OV motion states within the analyzed sample. The second hypothesis (H2) was not statistically supported at the 0.05 level. However, the descriptive heatmap suggested exploratory clustering between longitudinal/lane-changing motion patterns and rear-end or sideswipe crash outcomes. The findings suggest that, within the analyzed sample, AV-related collisions tend to concentrate around a limited number of recurrent and observable motion-interaction configurations. These patterns are interpreted as empirical co-occurrence structures rather than causal evidence of validated behavioral mechanisms. The proposed IBCT framework provides a reproducible basis for organizing and interpreting AV crash reports according to observable pre-crash motion configurations. While the findings may inform future safety monitoring and scenario-based assessment, they should not be read as direct policy recommendations.