From MIRAGE to CLEAR: Component-Level Explainable Anomaly Reasoning for Autonomous Vehicle Perception Systems

Abstract: Autonomous vehicle (AV) perception systems remain vulnerable to failures that current anomaly detectors can flag but cannot trace to a specific component; an attribution gap that impedes forensics and emerging transparency mandates like the EU AI Act. We introduce a unified framework comprising MIRAGE and CLEAR. MIRAGE (MUTCD-Informed Realistic Anomalous Generation Engine) integrates 48,022 real-world driving scenes from nuScenes, Waymo Open, and Argoverse 2 with structured rules from the Manual on Uniform Traffic Control Devices (MUTCD), generating 5,847 annotated anomalous scenarios with known module-level ground truth across four violation types: direct, subtle, contextual, and environmental. This data powers CLEAR, a hierarchical three-layer LLM pipeline that detects anomalies, classifies violations, and attributes failures to Traffic Sign Recognition (TSR), Automated Lane Centering (ALC), or Object Detection (OD) with interpretable justifications. Confidence-gated propagation and schema-constrained outputs prevent error cascades and minimize hallucinations. CLEAR achieves 95.2% detection accuracy, 62.3% classification accuracy, and 84% attribution accuracy on direct TSR violations. A Top- 2 evaluation paired with per-module confidence analysis raises overall attribution to 74.6%, confirming that apparent misattributions largely reflect the multi-module nature of anomaly types rather than reasoning failures. These results show that grounding LLM reasoning in structured traffic regulations enables reliable, interpretable forensics for AV perception systems, offering a practical path toward auditable, regulation-compliant safety analysis.