Enhancing Autonomous Vehicle Decision-Making Through Policy Transfer With Large Language Model

The safety and adaptability of decision-making systems to dynamic traffic scenarios are of paramount importance in the advancement of autonomous vehicles. Recent developments in generative artificial intelligence have shown promising effectiveness in scene understanding tasks within the context of traffic environments. Building on the insights gained from large language models for scene understanding, this paper introduces a novel transfer decision framework termed Adaptive Scenario Risk-based Reinforcement Learning (ASR-RL). This framework draws inspiration from human drivers who utilize their experiential knowledge when faced with new situations and consists of three key steps. First, we construct a task Markov decision process model based on various driving scenarios to pre-train policies across a series of small-scale scenarios, thereby creating a foundational policy library. Next, we introduce an adaptive scenario risk decision-making architecture that uses large language models to navigate the complexities of task vector resolution in unfamiliar environments. Finally, we transfer the driving policies from our policy library to new scenarios, enabling adaptive decision-making customized to specific driving contexts. Simulation experimental tests in three extended driving scenarios demonstrate that the proposed method guarantees safety and task completion under untrained conditions, achieving overall decision-making performance superior to mainstream methods, with performance in certain scenarios comparable to that of specialized training approaches. This work contributes to the development of autonomous driving decision-making systems with broad scenario adaptability.