Automated Evaluation of Large Vision-Language Models on Self-driving Corner Cases