Robust Vehicle Control with Smoothing and Prediction under Delayed Communication

Deep reinforcement learning (DRL) based vehicle controllers have shown excellent performance in autonomous driving. While these controllers deliver high performance, their computational demands often require task offloading to external servers. Yet, communication delays can hinder this process, forcing reliance on a simpler local module and compromising driving efficiency. To address this, we propose a framework that integrates action smoothing and local action prediction to enhance performance while ensuring safety under offloading delays. By incorporating a smoothing term into the reinforcement learning reward and introducing a polynomial based smoothing and prediction method, our approach ensures smoother transitions and robust driving. Experiments in the CARLA simulator demonstrate improved vehicle speed, reduced center deviations, and higher track completion rates under imperfect communication conditions.