Active suspension preview control via deep reinforcement learning with hybrid synchronizing policy and hierarchical road perception

Abstract

The rapid advancement of vehicle electronic control systems has provided favorable support for enhancing the dynamic response and control performance of suspension systems. Conventional control strategies based on suspension state feedback are limited in their active adjustment capability due to the lack of utilization of forward-looking road information, thereby affecting the further improvement of ride comfort. Although preview control methods that incorporate road preview information have partially mitigated this issue, existing approaches still suffer from deficiencies in effectively representing and efficiently utilizing road information. To address these challenges, this paper proposes an integrated perception-decision control framework that incorporates road preview information into the decision-making process of a reinforcement learning policy, thereby further enhancing ride comfort via reinforcement learning-based preview control. Specifically, a hierarchical road perception method is developed, which fuses road images and point clouds to extract rich and comprehensive road features. Furthermore, a hybrid synchronizing policy algorithm that dynamically couples on-policy and off-policy learning paradigms is introduced to balance training stability and sample efficiency, thus enabling more effective exploitation of the enriched perceptual inputs. Under various representative road conditions, experimental results demonstrate that the proposed method consistently outperforms the advanced comparative baselines.