Deep reinforcement learning for autonomous driving decision-making: Algorithms, applications and challenges

Abstract

Autonomous driving has the potential to greatly enhance road safety, traffic efficiency, and overall transportation effectiveness. The decision-making module serves as the core intelligence of the system, producing high-level commands for lane keeping, lane changing, and adaptive cruise control. Deep reinforcement learning has emerged as a pivotal approach in this domain, owing to its capacity to learn driving strategies through environmental interaction. This paper presents a systematic review of decision-making methods based on deep reinforcement learning for autonomous vehicles. It starts by outlining the general architecture of autonomous driving systems and emphasizing the central role of the decision-making module. The paper then explains the theoretical foundations of deep reinforcement learning, along with their applicability in driving decision tasks. Through a range of common driving scenarios, including lane keeping, lane changing, overtaking, and intersection navigation, the paper analyzes the decision-making advantages of deep reinforcement learning in complex and dynamic settings. Furthermore, it addresses key challenges in applying deep reinforcement learning to driving tasks. Compared to existing reviews, this study provides a systematic analysis of deep reinforcement learning in multiple decision-making scenarios for autonomous driving, overcoming the limitation of previous reviews that often focus on single scenarios. Adopting a progressive narrative structure, from single lateral or longitudinal control to increasingly complex tasks involving lateral-longitudinal coordination, it systematically clarifies the characteristics of different deep reinforcement learning algorithms and their applicable conditions. This structured approach helps researchers select appropriate algorithms and fusion strategies when facing driving scenarios of varying complexity.