RoadBench is a comprehensive benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on fine-grained spatial understanding and reasoning tasks specifically in urban road scenarios. The benchmark covers diverse driving conditions including intersections, parking lots, and complex road layouts to assess models' capabilities in spatial perception, object localization, and scene comprehension. It provides standardized evaluation metrics and extensive testing scenarios to advance the development of AI systems for autonomous driving applications. The benchmark includes both perception-level and reasoning-level tasks to comprehensively measure the performance of multimodal models in understanding complex road environments.