As AVs grow in complexity and diversity, identifying the root causes of operational failures has become increasingly complex. The heterogeneity of system architectures across manufacturers, ranging from end-to-end to modular designs, together with variations in algorithms and integration strategies, limits the standardization of incident investigations and hinders systematic safety analysis. This work examines real-world AV incidents reported in the NHTSA database. We curate a dataset of 2,168 cases reported between 2021 and 2025, representing more than 80 million miles driven. To process this data, we introduce CRASH, Cognitive Reasoning Agent for Safety Hazards, an LLM-based agent that automates reasoning over crash reports by leveraging both standardized fields and unstructured narrative descriptions. CRASH operates on a unified representation of each incident to generate concise summaries, attribute a primary cause, and assess whether the AV materially contributed to the event. Our findings show that (1) CRASH attributes 64% of incidents to perception or planning failures, underscoring the importance of reasoning-based analysis for accurate fault attribution; and (2) approximately 50% of reported incidents involve rear-end collisions, highlighting a persistent and unresolved challenge in autonomous driving deployment. We further validate CRASH with five domain experts, achieving 86% accuracy in attributing AV system failures. Overall, CRASH demonstrates strong potential as a scalable and interpretable tool for automated crash analysis, providing actionable insights to support safety research and the continued development of autonomous driving systems.
Figures & tables
Work
Stats
Expert Rules
NLP/LLM
Scalable
Sys. Reason
Favaro et al. [ 7 ]
✓
–
–
–
–
Houseal et al. [ 11 ]
✓
–
–
–
–
Shah et al. [ 21 ]
–
✓
–
–
✓
Zhang et al. [ 27 ]
✓
–
✓
✓
–
CRASH (Ours)
✓
✓
✓
✓
✓
Table 1: Methodological comparison of selected prior work and CRASH.
Figure 1: CRASH Architecture. During preprocessing, the database is filtered for incomplete entries and unified into four columns. Then, each row is sent to Processing through the LLM. We finally aggregate all information into a CSV file and send it to Postprocessing, where the data is used to generate our analysis and simulation descriptions.
Model
Make
Cases
%
Jaguar I-Pace
Waymo LLC
1015
46.83%
Cruise Av
GM LLC / Cruise LLC
355
16.37%
Jaguar I-Pace
Transdev
172
7.93%
Honda Civic
Honda
113
5.21%
Toyota Highlander
Zoox, Inc.
88
4.06%
Others
Various
425
19.60%
Table 2: Distribution of Incident Cases by Reporting Entity
Dataset Category
Original Cases
Final Cases
ADS
1790
1764
ADAS
2582
352
Others
3520
52
Total
7892
2168
Table 3: Case distribution before and after filtering
Figure 2: System prompt for the CRASH Agent, incorporating heuristic rules and one-shot examples.
Figure 3: Compact taxonomy of AV incident causes.
Method
AV Fail
Late AI
Cause
Sys. Fail
Majority class
54%
34%
42%
22%
Keyword rules
48%
46%
44%
18%
CRASH (ours)
86%
84%
76%
46%
Table 4: Baseline comparison on the 50 expert-evaluated cases (lenient criterion).
Figure 4: Primary cause distribution and subsystem breakdown across 2,168 AV incidents.
Figure 5: Overlap between late AI behavior, AV failures, and rear-end collisions ( N=2,168 ).
The Hong Kong University of Science and Technology (Guangzhou), China · Institute of Automation Chinese Academy of Sciences, Beijing, China · The University of Hong Kong, Hong Kong, China +1