An examination of real-world disengagement patterns of automated driving systems in autonomous-mode: A deep-learning...

Widespread adoption of automated driving systems (ADS) depends fundamentally on their operational reliability. Frequent or unexpected disengagements pose a significant barrier by undermining user trust. Although human-automation interaction has been studied, key gaps remain in understanding contemporary disengagement trends, their underlying causes, and the implications of disengagement type (ADS-initiated versus driver-initiated). This paper addresses these gaps to provide insights for future technical development and policy frameworks.

This analysis utilized 23,505 disengagement reports from the California DMV's Autonomous Vehicle Tester program, covering Automated Driving Systems (ADS) at SAE Levels 3-5 for the period 2019-2022. A deep learning-based natural language processing (NLP) pipeline was developed to extract causal factors from unstructured disengagement text descriptions, followed by manual validation to classify events into seven root-cause categories: Control, Planning, Perception, Software/Hardware, Localization/Mapping, Prediction, and Other Road Users. Temporal trends and associations between disengagement types and causal factors were evaluated using statistical analyses, including logistic regression.

Three key findings were identified. First, a significant temporal decline was observed in the proportion of disengagements initiated by the ADS (p < 0.01). Second, the root causes of disengagements exhibited a strong dependence on the initiator. Driver-initiated disengagements were predominantly associated with deficiencies in Perception, Planning, and Localization/Mapping. Conversely, ADS-initiated disengagements were overwhelmingly linked to Control-related failures. Finally, the failure profile of ADS-initiated disengagements evolved significantly between 2019 and 2022. Control-related issues, initially the dominant cause, became rare, while Software/Hardware issues emerged as the new leading cause, occurring at a rate three times that of driver-initiated disengagements by 2022.

While declining ADS-initiated disengagements signal improved system robustness, persistent failures in Control and emerging Software/Hardware issues indicate unresolved challenges. The divergent causes of driver- versus ADS-initiated events-suggesting preemptive human intervention versus forced system handovers-reveal distinct operational failure modes. The increasing prominence of Software/Hardware problems likely stems from growing system complexity. Consequently, targeted edge-case validation and adaptive trust calibration are critical for future ADS development.