Comfortable driving control for connected automated vehicles based on deep reinforcement learning and knowledge transfer

With the development of connected automated vehicles (CAVs), preview and large-scale road profile information detected by different vehicles become available for speed planning and active suspension control of CAVs to enhance ride comfort. Existing methods are not well adapted to rough pavements of different districts, where the distributions of road roughness are significantly different because of the traffic volume, maintenance, weather, etc. This study proposes a comfortable driving framework by coordinating speed planning and suspension control with knowledge transfer. Based on existing speed planning approaches, a deep reinforcement learning (DRL) algorithm is designed to learn comfortable suspension control strategies with preview road and speed information. Fine-tuning and lateral connection are adopted to transfer the learned knowledge for adaptability in different districts. DRL-based suspension control models are trained and transferred using real-world rough pavement data in districts of Shanghai, China. The experimental results show that the proposed control method increases vertical comfort by 41.10% on rough pavements, compared to model predictive control. The proposed framework is proven to be applicable to stochastic rough pavements for CAVs. This study proposes a comfortable driving framework by coordinating speed planning and suspension control with knowledge transfer. Based on existing speed planning approaches, a deep reinforcement learning (DRL) algorithm is designed to learn comfortable suspension control strategies with preview road and speed information. Fine-tuning and progressive networks are adopted to transfer the learned knowledge for adaptability in different districts. image