Title: TrafficClaw: A Generalizable LLM Agent in the Unified Physical Environment for Urban Traffic Control

Abstract:

While Large Language Model (LLM) agents have demonstrated significant proficiency in long-horizon reasoning, decision-making, and tool usage within digital settings, applying these capabilities to physically grounded systems presents substantial difficulties. In contrast to web, gaming, or coding environments—where goals are typically loosely connected—physical systems are characterized by tightly coupled dynamics. In these systems, local actions trigger cascading effects across interacting subsystems over time. Urban traffic control serves as a prime example of this complexity, as freeways, public transit networks, taxi services, and traffic signals continuously interact via shared spatial infrastructure and fluctuating temporal mobility demands. Current methods, including reinforcement learning (RL), traditional optimization, and LLM-based strategies, mostly target isolated subsystems, thereby restricting the potential for coordinated reasoning and holistic system optimization. To address this, we introduce TrafficClaw, a generalizable, LLM-driven agent designed for traffic control in physical urban contexts. TrafficClaw functions within a unified traffic framework that reveals interconnected urban dynamics and feedback loops. It employs executable spatiotemporal reasoning supported by persistent memory to facilitate long-term adaptation and utilizes multi-stage agentic RL to achieve coordinated, system-wide optimization. Our experimental results, spanning six traffic-control tasks across three major metropolitan areas, highlight the agent’s robustness, strong generalization capabilities, and effectiveness in cross-subsystem coordination. The project repository can be accessed at https://github.com/usail-hkust/TrafficClaw.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC