Docking assistance method for autonomous berthing by backward-time imitation learning and kernel density estimation based on AIS data

Autonomous berthing is desirable to assist or replace human operators to reduce the burden on navigation officers and resolve labor shortages. Over the past few decades, while researchers have proposed various methods to realize autonomous berthing, no online trajectory-planning method exists that reflects the intuition of skilled captains and is guaranteed to be constrained to terminal states. Additionally, most previous studies assumed a single reference trajectory, which tended to lack important information on the permissible deviation from the reference trajectory. This study presents a docking assistance method that integrates backward-time imitation learning (BTIL) and kernel density estimation (KDE). BTIL enabled a computational agent to generate captain-like docking trajectories without tuning the objective functions while imposing implicit constraints on terminal states. The proposed method offered trajectory and guidance distributions by applying multidimensional KDE to the BTIL results. The trajectory distributions indicated the area through which the ship should pass for successful berthing, whereas the guidance distributions indicated the velocities the ship should maintain at arbitrary locations. The performance of the docking assistance method was verified using Automatic Identification System (AIS) data from Shinmoji Port in Fukuoka Prefecture, Japan. • Docking assistance method for autonomous berthing was presented. • Backward-time imitation learning contributed to generate captains-like trajectories. • Trajectory distributions suggested in which area the ship should pass. • Guidance distributions indicated what velocities the ship should take. • The distributions were verified using actual trajectories obtained from AIS data.