Abstract
Recent advancements in autonomous vehicles (AVs) use Large Language Models(LLMs) to perform well in normal driving scenarios. However, ensuring safety indynamic, high-risk environments and managing safety-critical long-tail eventsremain significant challenges. To address these issues, we propose SafeDrive, aknowledge- and data-driven risk-sensitive decision-making framework to enhanceAV safety and adaptability. The proposed framework introduces a modular systemcomprising: (1) a Risk Module for quantifying multi-factor coupled risksinvolving driver, vehicle, and road interactions; (2) a Memory Module forstoring and retrieving typical scenarios to improve adaptability; (3) aLLM-powered Reasoning Module for context-aware safety decision-making; and (4)a Reflection Module for refining decisions through iterative learning. Byintegrating knowledge-driven insights with adaptive learning mechanisms, theframework ensures robust decision-making under uncertain conditions. Extensiveevaluations on real-world traffic datasets, including highways (HighD),intersections (InD), and roundabouts (RounD), validate the framework's abilityto enhance decision-making safety (achieving a 100% safety rate), replicatehuman-like driving behaviors (with decision alignment exceeding 85%), and adapteffectively to unpredictable scenarios. SafeDrive establishes a novel paradigmfor integrating knowledge- and data-driven methods, highlighting significantpotential to improve safety and adaptability of autonomous driving in high-risktraffic scenarios. Project Page: https://mezzi33.github.io/SafeDrive/