ConvoyLLM: Dynamic Multi-Lane Convoy Control Using LLMs

TL;DR

ConvoyLLM introduces a per-vehicle LLM-based decision framework augmented by a locally dynamic distributed graph to enable adaptive, stable multi-lane convoy formation. The approach integrates a reasoning module, a shared memory system, and a trajectory planning module to handle obstacle avoidance, joining, leaving, and escort formation switching. Validation in SUMO across diverse traffic densities demonstrates robustness and adaptability, with performance metrics indicating favorable safety and efficiency trade-offs. The work highlights practical potential for intelligent connected vehicle coordination and provides code at the project's GitHub repository.

Abstract

This paper proposes a novel method for multi-lane convoy formation control that uses large language models (LLMs) to tackle coordination challenges in dynamic highway environments. Each connected and autonomous vehicle in the convoy uses a knowledge-driven approach to make real-time adaptive decisions based on various scenarios. Our method enables vehicles to dynamically perform tasks, including obstacle avoidance, convoy joining/leaving, and escort formation switching, all while maintaining the overall convoy structure. We design a Interlaced formation control strategy based on locally dynamic distributed graphs, ensuring the convoy remains stable and flexible. We conduct extensive experiments in the SUMO simulation platform across multiple traffic scenarios, and the results demonstrate that the proposed method is effective, robust, and adaptable to dynamic environments. The code is available at: https://github.com/chuduanfeng/ConvoyLLM.

Figures (11)

• Figure 1: The overall framework of the multi-lane convoy formation control method. It contains a total of five modules: environment, reasoning, shared memory, trajectory planning, and control. The reasoning module obtains the perception results from the environment and generates the target lanes and target speeds of the vehicles, the trajectory planning module obtains the target values and generates the trajectories of each vehicle in the convoy, and finally the control module outputs the acceleration and steering angle commands, which are then applied to the environment.
• Figure 2: A case of the Reasoning module process. This simple obstacle avoidance scenario illustrates how the reasoning module collects information from the ego vehicle, environment vehicles, and neighbors, then generates a scene description for decision-making by the large model. In the figure, veh7 changes lanes to the right due to a slow vehicle ahead, while veh3 outputs an IDLE decision to follow the neighboring vehicle in the same lane.
• Figure 3: Interlaced formation and neighboring nodes diagram
• Figure 4: Avoiding obstacles scenario
• Figure 5: Average speed distribution of convoy under different traffic densities