Mental State Adaptive Takeover Warnings for Automated Driving Using a RAG-Enabled Large Language Model

Conditionally automated vehicles may require drivers to resume control within seconds, but takeover warnings may not account for drivers’ mental states. This study developed and evaluated a retrieval-augmented generation (RAG) system that assigned mental-state categories to synthetic profiles and generated profile-tailored takeover warnings. A search of four databases identified 6,049 records, of which 150 publications formed the retrieval corpus. The system was tested with 24 profiles representing anger, sadness, happiness, fatigue, mind-wandering, and external distraction. For 20 profiles (83.3%), the assigned category met the prespecified criterion by matching the intended category or a prespecified acceptable alternative. Fourteen participants rated the messages. Overall perceived helpfulness averaged 4.86 on a 7-point scale. Ratings for anger, sadness, mind-wandering, and fatigue were significantly above the scale midpoint. The findings provide preliminary evidence that RAG can generate profile-tailored candidate warnings from synthetic profiles. Simulator studies using measured driver data are needed to evaluate takeover behavior.