Abstract
Autonomous driving technology holds the promise of transforming transportation by enhancing road safety, reducing congestion, and improving mobility. However, achieving reliable and safe autonomous highway driving presents unresolved challenges, including accurately predicting the intentions of surrounding vehicles, making decisions under uncertainty, interpreting the rationale behind those decisions, and achieving human-like driving behaviour. Addressing these issues is crucial for the seamless integration of autonomous vehicles (AVs) into real-world traffic.
This thesis investigates the current problems to step forward in autonomous highway driving, focusing on three core challenges: prediction-based decision-making, human-like driving behaviour modelling, and interpreting decision-making processes. The research introduces novel methods to incorporate the intentions of surrounding vehicles into lane-change decisions, leveraging reinforcement learning (RL) to ensure adaptability and generalizability. To mitigate RL’s inefficiencies, such as its reliance on reward engineering and susceptibility to long-term rewardbased exploration, we propose alternative methods inspired by imitation learning to model human-like behaviour. Additionally, this work explores the explainability of decision-making processes, addressing the "black-box" nature of neural networks often associated with RL. By combining RL with Proportional-Integral-Derivative (PID) control and Large Language Models (LLMs), we propose HighwayLLM, an innovative framework that enhances decision-making through interpretability and reasoning capabilities. The thesis also emphasizes the transition from simulationbased algorithms to real-world applications, validating proposed methods on scaled research vehicles to account for sensor inaccuracies, environmental noise, and computational constraints. The findings demonstrate that the proposed methods significantly improve the safety, reliability, and adaptability of AVs in dense highway driving scenarios. By addressing critical challenges such as interpretability, decisionmaking under uncertainty, and human-like behaviour modelling, this research contributes to advancing autonomous highway driving, offering novel insights for future applications.