Speak to Simulate: An LLM-Guided Agentic Framework for Traffic Simulation in SUMO
Minwoo Jeong, Jeeyun Chang, Yoonjin Yoon
Traffic simulation plays a pivotal role in shaping transportation policy amid increasing urban traffic congestion and evolving mobility patterns. As a leading simulator, Simulation of Urban MObility (SUMO) is a widely adopted open-source platform for urban mobility research. However, generating traffic scenarios in SUMO is a time-consuming and intricate process, creating a steep learning curve and restricting iterative policy experiments. Although recent studies have integrated Large Language Models (LLMs) to mitigate these challenges, they often rely on fixed pipelines that lack interactive scenario generation and broad coverage of SUMO’s capabilities. To address these limitations, we propose AgentSUMO, an LLM-guided agentic framework for traffic simulation. Our framework positions the LLM as an AI agent, enabling interactive scenario design and flexible policy experimentation across the full range of SUMO functionalities. On the Manhattan network, our framework demonstrates that AgentSUMO addresses high-level policy goals through task planning and policy selection, resulting in consistent improvements in key traffic flow metrics. This novel agentic approach significantly lowers the barrier to entry for traffic simulation.