A review of reinforcement learning-based decision-making for autonomous driving in risk scenarios

Purpose Risk scenarios are a primary source of decision failures in autonomous driving. This study aims to systematically review reinforcement learning–based decision-making in such scenarios, analyze key challenges and future research directions and provide theoretical and methodological support for the development of safe and reliable autonomous driving systems. Design/methodology/approach This study analyzes the characteristics and decision-making complexity of risk scenarios, and summarizes reinforcement learning frameworks and core methods for addressing these challenges. It focuses on three representative scenario types and examines their impacts on state representation, reward design and decision-making mechanisms. To address data scarcity, recent advances in safety-critical scenario generation are reviewed, comparing data-driven, adversarial and knowledge-based approaches and summarizing their key features and development trends. Findings The results show that different risk scenarios require distinct environment modeling and decision-making mechanisms, placing higher demands on state representation, reward design and policy learning in reinforcement learning. Although safety-critical scenario generation methods help mitigate data scarcity, trade-offs among realism, diversity and controllability remain. Key challenges persist in dynamic risk modeling, human–vehicle interaction and safety verification and assurance. Originality/value This study systematically integrates existing research on reinforcement learning for autonomous driving decision-making from the perspective of risk scenarios, and establishes a comprehensive analytical framework covering scenario modeling, decision-making methods and data generation. It further provides an in-depth discussion of emerging issues, offering valuable insights and directions for future research.