Autonomous Ship Collision Avoidance Trained on Observational Data

Raphael Schwinger, Ghassan Al-Falouji, Sven Tomforde

Abstract: Marine Autonomous Surface Ships (MASS) are gaining interest worldwide with the potential to reshape mobility and freight transport at sea. Collision avoidance and path planning are central components of the intelligence of a MASS. While Deep Reinforcement Learning (DRL) techniques often learn these abilities in a simulated environment, this article explores an alternative approach: learning collision avoidance and path planning solely from observational data, thus minimizing the need for simulator-based training. A state-action dataset of ship trajectories is constructed from recorded Automatic Identification System (AIS) messages. Using this data, we examine the application of the Prediction and Policy-learning Under Uncertainty (PPUU) technique, which involves training an action-conditional forward model and learning a policy network by unrolling future states and back-propagating errors from a self-defined cost function. To evaluate the learned policy, FerryGym, a Gymnasium environment is developed for evaluating the policy network using observational data.

External IDs:dblp:conf/arcs/SchwingerAT23