Using Retrieval-Augmented Generation to Support Parameter-Specific Takeover Request Personalization in Automated Driving

Conditionally automated vehicles still require drivers to retake control when automation reaches its operational limits, so takeover request (TOR) design remains a safety concern. This study developed and evaluated a retrieval-augmented generation (RAG) system that uses a large language model (LLM) to produce parameter-specific, literature-grounded TOR plans. Two literature corpora on takeover factors and alert modality served as the knowledge base, and a structured situation object represented driver traits, driver state, non-driving-related tasks, and driving scenario conditions. The system generated TOR plans across 10 parameter categories for 40 persona-scenario combinations. Several TOR parameters varied with takeover-relevant inputs in the generated test set, and modality recommendations changed according to the sensory impairments described in the situation object. These preliminary results suggest that RAG-based LLMs may support parameter-specific TOR personalization in automated driving.