清华最新ADRD：自动驾驶决策树模型实现可解释性与性能双突破！

Core Viewpoint - The article discusses the rapid advancements in the autonomous driving field, emphasizing the increasing demand for transparency and interpretability in decision-making modules of autonomous systems. It highlights the limitations of both data-driven and rule-based decision systems and introduces a novel framework called ADRD, which leverages large language models (LLMs) to enhance decision-making capabilities in autonomous driving [1][2][26]. Summary by Sections 1. Introduction - The autonomous driving sector has seen significant progress, leading to a heightened focus on the interpretability of decision-making processes within these systems. The reliance on deep learning methods has raised concerns regarding performance in non-distributed driving scenarios and the complexity of decision logic [1]. 2. Proposed Framework - The ADRD framework is introduced as a solution to the challenges faced by traditional decision systems. It combines rule-based decision-making with the capabilities of LLMs, demonstrating superior performance in various driving scenarios compared to conventional methods [2][26]. 3. Algorithm Model and Implementation Details - The ADRD model consists of three main modules: information, agent, and testing. The information module converts driving rules and environmental data into natural language for LLM processing. The agent module includes a planner, encoder, and summarizer, which work together to ensure stable reasoning and effective feedback loops [5][7][13]. 4. Experimental Results - Experiments conducted in the Highway-env simulation environment show that ADRD outperforms traditional methods in terms of average safe driving time and reasoning speed across various driving conditions. For instance, in a normal density scenario, ADRD achieved an average driving time of 25.15 seconds, significantly higher than other methods [21][22]. 5. Conclusion - The article concludes that the ADRD framework effectively utilizes LLMs to generate decision trees for autonomous driving, outperforming both traditional reinforcement learning and knowledge-driven models in performance, response speed, and interpretability [26].