LLM-Powered Agent for Dense Pedestrian Flow Path Planning in Train Stations

Wenshuai Xu, Xiaoyan Wang, Zijin Liu, Xiang Zhao, Zhiyong Cui, Xu Geng, Zhoujin Tang, You Song

Keywords: LLM-powered Agent, Crowd Simulation, Congestion Analysis, Destination Recommendation, Path Planning

Abstract: Managing dense pedestrian flows in train stations is a critical yet challenging task in urban transportation. Although LLM-powered agents show strong decision-making ability in many domains, they struggle in crowded, dynamic environments due to the scarcity of high-quality spatiotemporal data and effective representation methods. These limitations often cause hallucinations and poor planning. Key challenges lie in unifying macroscopic crowd metrics with microscopic individual behaviors and in efficiently encoding fine-grained spatiotemporal data under LLM context constraints. To overcome these issues, we propose $\mathbf{SP^3Agent}$, an LLM-powered agent for pedestrian flow planning that leverages a simulator, structured knowledge augmentation, and dedicated computational and analytical tools. Our simulator generates macro-scale density and velocity distributions from micro-scale trajectories, enabling holistic scene understanding. The knowledge augmentation leverages the KG-RAG framework to effectively retrieve and represent relevant spatiotemporal knowledge, while dedicated tools—such as the congestion analytics module—enable real-time, on-demand analysis of crowd dynamics. Extensive evaluations conducted in a high-fidelity environment simulated from Beijing West Station demonstrate that our agent significantly improves evacuation efficiency. Compared to conventional simulation-only crowd dynamics, our LLM agent achieves a 62\% reduction in total evacuation time, a 50\% decrease in average time cost, and a 21\% shortening of average path length. These results demonstrate the potential of our approach in leveraging simulation-augmented data to mitigate LLM hallucination in numerical-intensive spatiotemporal decision tasks, offering a robust framework for real-world deployment in transportation hubs.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 3567