Large language model guided deep reinforcement learning for safe autonomous vehicle decision making

Deep reinforcement learning (DRL) has shown promising potential for decision-making in autonomous driving. However, it requires extensive interaction with the environment and generally has low learning efficiency. To address these challenges, this paper proposes a novel large language model (LLM) guided deep reinforcement learning (LGDRL) framework for the decision-making problem in autonomous driving. Leveraging the powerful reasoning capabilities of LLMs, an LLM-based driving expert is designed to provide intelligent guidance in the DRL learning process. Subsequently, an innovative expert policy constrained algorithm and a novel LLM-intervened interaction mechanism are developed to efficiently integrate the guidance from the LLM expert to enhance the performance of DRL decision-making policies. Extensive experiments are conducted to evaluate the performance of the proposed LGDRL method. The results demonstrate that our proposed method effectively leverages expert guidance to enhance both learning efficiency and performance of DRL, achieving superior driving performance. Moreover, it enables the DRL agent to maintain consistent and reliable performance in the absence of LLM expert guidance, which is promising for real-world applications. The supplementary videos are available athttps://bitmobility.github.io/LGDRL/.