This paper presents RoadBench, a comprehensive benchmark designed to evaluate Multi-Modal Large Language Models (MLLMs) on fine-grained spatial understanding and reasoning tasks in urban road scenarios. The benchmark includes diverse urban road images, detailed annotations, and challenging questions that require precise spatial perception and reasoning. RoadBench aims to address the gap in existing benchmarks that lack focus on fine-grained spatial understanding and reasoning under complex urban road conditions. Experimental results demonstrate that current MLLMs still face significant challenges in handling fine-grained spatial understanding and reasoning in urban road scenarios.