Receding-Horizon Multi-Kernel Reinforcement Learning for Motion Planning of Intelligent Vehicles

Optimal motion planning is critical to improving the safety, maneuverability, and comfort of intelligent vehicles. However, existing optimization-based methods struggle to simultaneously handle three key challenges: 1) unknown nonlinear vehicle dynamics that invalidate conventional model-based control; 2) time-varying environmental constraints like obstacle dynamics, road geometry variations, and perception uncertainty requiring online adaptability; and 3) real-time computational demands for dynamic traffic scenarios. This article proposes a receding-horizon multi-kernel reinforcement learning for motion planning (RMRL-MP) approach that addresses the optimal motion planning problem with unknown nonlinear dynamics and time-varying environment constraints. First, RMRL-MP adopts the receding-horizon reinforcement learning framework, which utilizes the actor–critic structure to learn the optimal motion planning policy in decomposed prediction horizons to efficiently solve nonlinear constrained problems in real time. Second, to further enhance the learning efficiency, the proposed method introduces multi-kernel features to represent the actor–critic structure and the dynamics model. Finally, real-world experiments and extensive simulation studies in various traffic scenarios demonstrate the effectiveness, feasibility, and superiority of the proposed framework.