Contextual-Personalized Adaptive Cruise Control via Fine-Tuned Large Language Models

Abstract

Adaptive cruise control (ACC) is a widely adopted technique within advanced driver assistance systems (ADAS) to alleviate driver workload and fatigue in long-distance driving or stop-and-go traffic scenarios. However, conventional ACC systems typically fail to account for drivers' preferences or changing environmental conditions, limiting their adaptability in adjusting headway. To bridge this gap, this study introduces a novel contextual-personalized ACC (CP-ACC) framework that leverages the contextual reasoning and adaptive customization potential of large language models (LLMs). Specifically, LLMs including LLaMA-3-8B and Mistral-7B are fine-tuned with a synthetically generated dataset encompassing diverse drivers' preferences (e.g., energy efficiency, comfort) and real-time contextual information (e.g., weather, traffic conditions). CP-ACC demonstrates the ability to identify, quantify, and balance competing objectives (e.g., safety and mobility) compared to linear feedback ACC and the intelligent driver model (IDM). Supervised fine-tuning (SFT) further enhances the LLMs' ability to recognize driving objectives and generate safe, context-aware longitudinal control commands, outperforming zero-shot and few-shot prompting. Overall, the proposed CP-ACC framework presents a promising direction for delivering smart, adaptive, and personalized driving assistance tailored to varying drivers' preferences and dynamic traffic environments.