Policy Graphs and Theory of Mind for Explainable Autonomous Driving

Autonomous driving has seen significant advancements over the past two decades, thanks to progress in artificial intelligence (AI). However, the decision-making processes of autonomous vehicles (AVs) remain largely opaque, creating a barrier to societal trust and regulatory acceptance due to concerns over trustworthiness, safety and accountability. This thesis explores the application of a novel explainable AI (XAI) technique, Policy Graphs (PGs), in the domain of autonomous driving. Policy Graphs represent an agent's policy as a directed graph with natural language descriptors, providing a human-readable explanation of the agent’s behaviour. This framework is further enhanced by incorporating notions of Theory of Mind, to understand these systems as if they possessed beliefs, desires and intentions, and enabling the graph to capture not just what the agent does, but also what it desires and intends to do. In particular, we make the following contributions. First, we provide a comprehensive review of current XAI techniques in the context of autonomous driving. Second, we develop a framework that integrates the agent’s desires and intentions into the Policy Graphs, to extract the intentions behind specific driving decisions and identify anomalous or undesirable behaviours. Finally, we explore how external factors, such as weather and lighting conditions, influence decision-making, uncovering potential biases and patterns under different driving conditions. The results demonstrate that combining Policy Graphs with concepts from Theory of Mind offer an effective approach to explaining and interpreting vehicle behavior, thereby improving the understanding of autonomous driving systems.