CRASH: Cognitive Reasoning Agent for Safety Hazards in Autonomous Driving

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.

Savvy: Trustworthy Autonomous Vehicles Architecture

The increasing interest in Autonomous Vehicles (AV) is notable due to business, safety, and performance reasons. While there is salient success in recent AV architectures, hinging on the advancements in AI models, there is a growing number of fatal incidents that impedes full AVs from going mainstream. This calls for the need to revisit the fundamentals of building safety-critical AV architectures. However, this direction should not deter leveraging the power of AI. To this end, we propose Savvy, a new trustworthy intelligent AV architecture that achieves the best of both worlds. Savvy makes a clear separation between the control plane and the data plane to guarantee the safety-first principles. The former assume control to ensure safety using design-time defined rules, while launching the latter for optimizing decisions as much as possible within safety time-bounds. This is achieved through guided Time-aware predictive quality degradation (TPQD): using dynamic ML models that can be tuned to provide either richer or faster outputs based on the available safety time bounds. For instance, Savvy allows to safely identify an elephant as an obstacle (a mere object) the earliest possible, rather than optimally recognizing it as an elephant when it is too late. This position paper presents the Savvy's motivations and concept, whereas empirical evaluation is a work in progress.