E2E neural predictive control for vessel: Autonomous berthing via DRL and PIM-MPC

In complex marine environments, vessel berthing operations face challenges including nonlinear dynamic effects, strong coupling issues, environmental uncertainties, underactuated characteristics, and multiple constraints. This study proposes a fully data-driven end-to-end vessel berthing control strategy based on Deep Reinforcement Learning (DRL). The study trained a Long Short-Term Memory (LSTM) neural network predictive model using historical vessel input–output data and designed a Model Predictive Controller (MPC) to achieve precise berthing operations. Furthermore, the study introduced a Random Forest model to enhance berthing accuracy and reduce control input fluctuations. To further improve berthing safety and real-time performance, the study developed a Refined Integration Method (RIM), proposing the PIM-MPC approach to optimize berthing maneuvers, with comparative analysis against the Random Forest model. Simulation results confirm the proposed method achieves high-precision berthing under complex environmental disturbances without requiring accurate vessel dynamics models. This approach enhances real-time performance and significantly reduces control input fluctuations compared to Random Forest methods, while simultaneously improving berthing safety despite requiring no prior berthing control input data. This method demonstrated its capability to achieve highly precise berthing of large vessels within confined port environments. Comprehensive full-scale vessel experiments rigorously validated its feasibility and effectiveness.