Pang, Hao (Beijing Institute of Tecnology), Wang, Zhenpo (Beijing Institute of Technology), li, guoqiang (beijing insititute of technology)

LLM-Driven Constrained MARL for Collaborative Multi-Vehicle Control in Autonomous Cooperative Transportation Systems

Scheduled for presentation during the Invited Session "S22c-Emerging Trends in AV Research" (TH-LA-T22), Thursday, November 20, 2025, 16:00−16:20, Coolangata 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Cooperative transportation systems (CTS) for heavy-duty payloads face significant challenges in multi-vehicle coordination due to complex coupling constraints. This paper proposes a novel Large Language Model (LLM)-Driven Constrained Multi-agent Reinforcement Learning (LDC-MARL) framework for collaborative control of CTS. To realize safe and stable multi-vehicle cooperative transportation, our approach integrates LLM guidance and fixed inter-vehicle distance constraints into the MARL policy optimization process and resolves the constrained MARL problem via Lagrangian duality theory. Extensive experiments show that LDC-MARL achieves superior performance, attaining a 100% success rate in task completion while significantly improving constraint satisfaction with up to 92.35% reduction in violations compared to state-of-the-art baselines. The results demonstrate the proposed framework’s effectiveness in developing collision-free and motion-coordinated CTS, improving its practical applicability. The supplementary videos are available at https://bitmobility.github.io/LDC-MARL/.