Enhancing safety in autonomous driving through reinforcement learning: A comparative study of machine learning approaches

Abstract

With the emergence of autonomous vehicles, the automotive industry promises to revolutionize human mobility, offering many advantages such as increased driving comfort, reduced congestion, and improved road safety. Despite significant advancements in sensor technology and perception algorithms, ensuring the safety of autonomous driving remains a critical challenge. This dissertation aims to explore how reinforcement learning techniques can be leveraged to further enhance the safety of autonomous driving systems. The study presents an in-depth review of the vast landscape of deep neural network and Reinforcement Learning (RL) methods, emphasizing their limitations in applicability and functional safety. Various modifications to state-of-the-art DQN RL algorithms are proposed, assessing their impact on training stability and agent performance. An essential contribution of this dissertation is the integration of profound safety body of knowledge, in alignment with automotive safety standards, with advanced machine learning expertise. This work extensively investigates the practical implementation of deep neural network classifiers, identifies safety risks inherent in every development phase, and puts forth both theoretical and practical solutions to address and mitigate these risks. A safety layer for RL agents, comprising eight key features, is introduced to enhance autonomous driving safety. This includes methods to quantify and optimize exploration behaviour in continuous state spaces. The safety layer integrates human expert guidance, prevents unsafe actions, imposes safety constraints, dynamically shapes rewards, introduces redundancy, and ensures a fail-safe strategy for Operational Design Domain (ODD) violations. An additional method enhances RL agent adaptability by emulating human drivers. In the final chapter, the study utilizes the OPEN AI Gym environment for highway driving experiments. Reinforcement learning-based agents are equipped with the safety layer features to make real-time decisions in dynamic and varied driving scenarios, and unexpected events. Quantitative comparisons of experimental results are drawn to assess RL agent performance, safety KPIs, and other relevant metrics. The findings of this dissertation contribute to the ongoing discourse on autonomous vehicles, offering valuable insights into the capabilities and limitations of RL-based autonomous driving systems. This work shall also increase awareness about the criticality of safety in AI-based solutions and guide how such sophisticated solutions comply with normative standards.