کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

RIDER: Residual Reinforcement Learning to Enhance Imitation-Based Autonomous Driving

نویسندگان :

Zahra Akhbariyoon¹ Samaneh Hosseini Semnani²

1- Isfahan University of Technology 2- Isfahan University of Technology

کلمات کلیدی :

Autonomous Driving،Residual Reinforcement Learning،Imitation Learning،Decision-Making،CALRA Simulator

چکیده :

— Urban autonomous driving faces significant challenges due to unpredictable behaviors, adverse weather conditions, and diverse types of roads. Traditional approaches towards imitation learning (IL) and reinforcement learning (RL) have exhibited complementary merits; however, both suffer from crucial shortcomings when employed individually. This paper introduces RIDER (Residual Inference Driving Enhancement Reinforcement), a novel hybrid architecture that leverages Residual reinforcement learning to enhance imitation-based systems for autonomous driving. The proposed method first pre-trains an IL model from expert demonstrations to establish a sound baseline policy and includes an RL component for fine-tuning, enabling the system to adapt to unseen scenarios by adding corrections to the base IL policy. This hybrid strategy significantly reduces the complexity of training and speeds up policy optimization compared to standalone RL while enhancing safety and efficiency in complex urban environments. Extensive experiments in the CARLA simulator show that the approach improves collision avoidance, lane-keeping, and route completion under diverse scenarios. Specifically, RIDER achieved a 4.5% improvement in success rate over the baseline IL model in test Town02 and on average 7.5% improvement in train Town01. These results underscore the potential of RIDER to enhance the safety and functionality of autonomous driving systems in real-world urban settings.