AgentsCoDriver: Large Language Model Empowered Collaborative Driving with Lifelong Learning

TL;DR

This work tackles interpretability, generalization, and lifelong learning in connected autonomous driving by introducing AgentsCoDriver, an LLM-powered framework that enables multi-vehicle collaboration and negotiation. It formulates the driving problem as a $D$-POMDP and integrates five modules—observation, cognitive memory, reasoning, reinforcement reflection, and communication—together with an evaluator/reflector loop to support lifelong learning. The approach achieves superior performance over baselines, demonstrates improved learning over time, and enables inter-vehicle negotiation to enhance safety and efficiency in complex traffic. These results highlight the practical potential of LLM-driven, memory-augmented, multi-agent collaboration for future autonomous driving systems.

Abstract

Connected and autonomous driving is developing rapidly in recent years. However, current autonomous driving systems, which are primarily based on data-driven approaches, exhibit deficiencies in interpretability, generalization, and continuing learning capabilities. In addition, the single-vehicle autonomous driving systems lack of the ability of collaboration and negotiation with other vehicles, which is crucial for the safety and efficiency of autonomous driving systems. In order to address these issues, we leverage large language models (LLMs) to develop a novel framework, AgentsCoDriver, to enable multiple vehicles to conduct collaborative driving. AgentsCoDriver consists of five modules: observation module, reasoning engine, cognitive memory module, reinforcement reflection module, and communication module. It can accumulate knowledge, lessons, and experiences over time by continuously interacting with the environment, thereby making itself capable of lifelong learning. In addition, by leveraging the communication module, different agents can exchange information and realize negotiation and collaboration in complex traffic environments. Extensive experiments are conducted and show the superiority of AgentsCoDriver.

Figures (9)

• Figure 1: The Scenario of Multi-Vehicle Negotiation.
• Figure 2: Overall Architecture of AgentsCoDriver. The architecture of AgentsCoDriver consists of five modules: observation module, reasoning engine, memory module, reinforcement reflection module, and communication module. The reasoning engine, communication module, and reinforcement reflection module leverage LLMs to generate messages and final decisions.
• Figure 3: Prompt design of the reasoning engine.
• Figure 4: Reinforcement Reflection Module.
• Figure 5: Visualization of two different scenarios. In Highway scenario, the ego-vehicle (The green one) is traveling on a highway with multiple lanes with various vehicles (The blue one). The goal of an agent is to achieve high velocity while ensuring it does not collide with nearby vehicles. In Intersection scenario, the task of the ego vehicle is to pass a busy intersection safely without any collision. The ego vehicle can turn left or right, or go straight.