Feedback-Enhanced Closed-loop Decision Making for Autonomous Driving with Large Language Models

Autonomous driving (AD) has great potential for enhancing traffic safety and transportation efficiency. However, conventional AD methods often lack sufficient generalization and reasoning capabilities in dynamic and complex scenarios. Although large language models (LLMs) can alleviate this limitation, their inherent hallucination problem may introduce safety risks into decision making. To address these challenges, this paper proposes FEDrive-LLM, a feedback-enhanced closed-loop autonomous driving framework. This framework integrates the core modules of environment interaction, memory retrieval, reasoning decision, online safety verification, and reflection feedback. Furthermore, the safety verification module in FEDrive-LLM conducts online collision risk assessment for driving decisions, effectively mitigating the safety risks caused by model hallucinations. Meanwhile, the reflection feedback module reviews potentially conservative decisions and generates concise suggestions for subsequent steps, helping the system achieve a better balance between safety and driving efficiency. Simulation results show that FEDrive-LLM significantly outperforms existing state-of-the-art methods across multiple metrics, including success rate, success steps, and average speed. These results demonstrate its robustness and effectiveness in complex traffic environments.