Evaluating LLM Driving Explanations through an Ego-Centric Scenario Taxonomy

For human--robot interaction in autonomous driving, understanding when and why automated systems can effectively explain their behavior is critical for transparency, trust, and user understanding. Large language models (LLMs) can generate natural-language explanations of driving scenes, yet it remains unclear whether some types of driving situations are inherently easier or harder for them to describe. To investigate this question, we introduce an ego-centric taxonomy of driving scenarios and apply it to the BDD-X test set, creating a category-aligned evaluation benchmark. Using this dataset, we compare the explanation performance of RAG-Driver and GPT-4o across both top-level and fine-grained scenario categories. For each model, explanation performance differs significantly across scenario categories, indicating that scenario type is a meaningful factor influencing explanation quality. These findings highlight the importance of scenario-aware evaluation when assessing explanation quality in autonomous driving.