BladeRunner-AD: Adversarial Co-evolution of Robust and Interpretable Driving Policies via LLM Agent Games

作者：Ao Guo, Jiachen Hou, Fei Wang · 年份：2025 · DOI：10.1109/mesa68091.2025.11278881 · 被引用次数：1 · 研究领域：Adversarial Robustness in Machine Learning、Explainable Artificial Intelligence (XAI)、Autonomous Vehicle Technology and Safety

The core challenge in autonomous driving decision-making systems lies in ensuring safety and robustness in complex and adversarial scenarios. Traditional reinforcement learning methods lack interpretability, while recent approaches that use Large Language Models (LLMs) to generate interpretable rules, though transparent, produce policies that are brittle against unknown adversarial corner cases. To address this issue, we propose BladeRunner-AD, an innovative framework for driving policy generation. This framework models the policy generation process as a zero-sum game between two LLM agents. A "Policy Defender" agent is responsible for generating an executable decision-tree driving policy, while a "stylized" Adversarial Attacker agent analyzes the current policy’s logical flaws and generates adversarial scenarios in simulation most likely to cause it to fail. Through a closed loop of generation-attack-reflection-revision, the policy co-evolves under continuous, worst-case pressure. This mechanism, akin to automated red teaming, forces the generated rules to be continuously refined, thereby achieving high robustness. We conducted comprehensive experiments in interactive scenarios on the highway-env simulation platform. The results demonstrate that policies generated by BladeRunner-AD exhibit a significantly lower collision rate on an independent adversarial test set compared to baseline methods, showcasing superior robustness and safety. This work presents a novel paradigm for...