Autonomous Driving Planning Based on Large Language Model: Collaborative Driving

Recent advancements in end-to-end autonomous driving have primarily relied on data-driven approaches. However, these methods face challenges regarding interpretability and safety guarantees in motion prediction results. We draw inspiration from the knowledge-driven paradigms used in human driving to address these issues. Our exploration focuses on how to incorporate similar capabilities into autonomous driving (AD) systems. We propose LLMDriver, a Large Language Model (LLM)-based agent that integrates an interactive driving environment, multiple driving agents, and a memory module. This framework leverages the potential of large language models (LLMs) to develop a system that drives like humans. By reasoning through the driving actions taken and accumulating experiences from continuous driving, our approach aims to enhance motion planning in autonomous vehicles. LLMDriver achieves an L2 error of 0.82 and a 0.31% collision rate in UniAD metrics, performing comparably to UniAD and slightly below RDA Driver. Under ST-P3, it attains a 0.41 L2 error and 0.11% collision rate. In CARLA closed-loop tests, it scores 65% in driving and achieves the highest route completion, surpassing prior methods by 1.1%.