Reinforcement learning-based NMPC for tracking control of ASVs: Theory and experiments

We present a reinforcement learning-based (RL) model predictive control (MPC) method for trajectory tracking of surface vessels. The proposed method uses an MPC controller in order to perform both trajectory tracking and control allocation in real-time, while simultaneously learning to optimize the closed loop performance by using RL and system identification (SYSID) in order to tune the controller parameters. The efficiency of the method is evaluated by performing simulations on the unmanned surface vehicle (USV) ReVolt, as well as simulations and sea trials on the autonomous urban passengers ferry milliAmpere. Our results demonstrate that the proposed method is able to outperform other state of the art methods both in tracking performance, as well as energy efficiency.