A Comprehensive Survey on Deep Reinforcement Learning for Autonomous Vehicles Urban Driving: Approaches, Challenges, and Future Directions

Deep Reinforcement Learning (DRL) integration has resulted in notable progress in the development of autonomous driving systems (ADS). DRL is an effective approach for autonomous vehicles (AVs) because of its capacity to manage intricate decision-making processes in dynamic and unpredictable contexts. This study provides a comprehensive review of DRL approaches applied to ADS, with a focus on key components such as planning, control, localization, mapping, and scene understanding. We examine various DRL methods: value-based, policy-based, and actor-critic, and highlight their contributions to real-time decision-making, low-level vehicle control, path planning, and trajectory optimization. In addition, we discuss advanced DRL extensions, such as imitation learning (IL), multi-agent systems, and safe DRL, which address important challenges related to safety, efficiency, and robustness in autonomous systems. This study also reviews modern simulation environments that play a crucial role in testing and training DRL models within virtual settings. Finally, we outline future research directions, emphasizing the need for improved generalization, sample efficiency, and ethical considerations in DRL-based ADS. The purpose of this survey is to provide insights that can contribute to the development of safer and more efficient AVs capable of operating under diverse and challenging conditions.