Autonomous Forklift Path Following Using Large Language Model Agents and Reinforcement Learning

This thesis explores the viability of utilising Large Language Models (LLMs) to iteratively tune automatic control parameters for an autonomous forklift truck performing a path-following task. Additionally, an established control policy learning method, Reinforcement Learning (RL), is implemented to perform the path-following task. The main objective is to evaluate the performance of each method, allowing for a comparison between them. Two systems are implemented to investigate the objective: a multi-agent LLM parameter tuning framework and a deep Q-network RL agent. The LLM framework focuses on tuning three control structures: Linear Quadratic (LQ), State Feedback, and Pure Pursuit, while examining the impact of different in-context learning prompt types, including zero-shot, zero-shot chain-of-thought, few-shot, and few-shot chain-of-thought prompting. The RL solution is trained in simulation to develop a policy that achieves desirable path-following performance. The results demonstrate the promising capability of the LLM-based framework to tune control parameters effectively using simulation feedback, achieving strong performance. Although the RL approach achieves good results when trained with a simplified system model, it struggles with policy transfer to the more realistic simulation environment. Both LLM and RL methods require significant time and effort to implement. The extensive in-context prompt design needed for consistent LLM tuning results is comparable in complexity to the reward function design and training required by the RL solution.