Abstract
As urbanisation accelerates, traditional Intelligent Transportation Systems (ITSs), reliant on manual monitoring and rule‐based logic, struggle to manage increasingly complex and dynamic traffic conditions. This study proposes TrafficRobot, a hierarchical AI agent framework that implements a closed‐loop control flow for urban traffic management. Built on a four‐layer architecture (Data, Foundation Model, Agent Platform, and Smart Application). TrafficRobot introduces several key innovations. First, it leverages large language models (LLMs) as decision‐making agents, enabling interpretable, human‐like reasoning through multi‐agent orchestration. Second, it develops a lightweight, domain‐specialised model, TrafficLLM (1.7B), using distillation and fine‐tuning to overcome terminology and efficiency limitations of general LLMs, allowing edge‐side deployment. Third, extensive experiments on real‐world datasets demonstrate that TrafficRobot achieves better performance compared to rule‐based and reinforcement learning methods in control metrics, with intrinsic interpretability through natural‐language reasoning. This work delivers a scalable and explainable solution for next‐generation ITS, offering practical pathways toward autonomous urban traffic intelligence.