Decoupled Deep Reinforcement Learning with Sensor Fusion and Imitation Learning for Autonomous Driving Optimization

Our study explores the integration of decoupled deep reinforcement learning (DRL) frameworks in autonomous driving, focusing on the impact of sensor fusion and imitation learning (IL) to enhance performance in complex, multi-task environments. Traditional DRL approaches often encounter limitations in real-time decision-making, particularly in the simultaneous handling of perception and control tasks. By decoupling these processes, the proposed method achieves improved specialization and adaptability, addressing the inherent challenges posed by dynamic driving scenarios. The comparative analysis demonstrates that the decDRL+IL+Early Fusion model reduces lane invasions by 79.6% and collision rates by 49.8% after 5,000 episodes, outperforming decoupled DRL models without IL, which yield reductions of 29.4% and 19.7%, respectively. Furthermore, early fusion improves task efficiency with a 14.7% reduction in steps per episode, while reducing energy consumption by 19.8%, underscoring its suitability for real-world deployment in autonomous systems. The findings emphasize the effectiveness of combining early sensor fusion with imitation learning, offering significant advancements in the scalability, efficiency, and safety of autonomous driving technologies.