Fine Tuning a Large Language Model for Tactical Decision Making in Level 3 Autonomous Trucks

Abstract

This thesis investigates whether a Large Language Model (LLM) can be adapted to serve as the tactical brain of a Level-3 autonomous truck through supervised fine-tuning (SFT). We first generated highway driving scenarios in the SUMO simulator, pairing each coded scenario with high-level maneuvering decisions, which include ACC set speed, time gap, lane change intent, generated by a powerful LLM. The resulting scenario-decision pairs constitute a domain-specific dataset that captures a variety of safety-critical interactions between a self-propelled truck and surrounding traffic. Three open-source modelsMeta-Llama-3.1-8B, Qwen 2.5-14B, and DeepSeek-R1-Distill-Llama-8B-are then fine-tuned with Low-Rank Adaptation (LoRA). A modular control stack separates the LLMs high-level reasoning from a low-level Intelligent Driver Model (IDM) that executes longitudinal and lateral motion, mirroring real-world practice.

Evaluation of SUMO episodes showed that fine-tuning improved the quality of decisions. All models improve the achieve a high success rate. Despite the fact that the fine-tuned LLMs achieved a high success rate, we discovered that the LLMs does not fully learn a perfect set of driving strategies. The LLMs does not completely learn the truck’s lane changing strategy. As a result, the LLMs behaved somewhat clumsily in some scenarios. After fine-tuning, some unsafe decisions were eliminated, which confirms the improvement of logical consistency. The models also generate concise natural language rationales, improving the interpretability and compliance of the system. This study shows that when equipped with a tailored driving dataset and efficient LoRA fine-tuning, a modestly sized LLM can provide a degree of safe, efficient, and interpretable but not perfect tactical decisions for self-driving trucks.