Dynamic Positioning using Deep Reinforcement Learning

This paper demonstrates the implementation and performance testing of a Deep Reinforcement Learning based control scheme used for Dynamic Positioning of a marine surface vessel. The control scheme encapsulated motion control and control allocation by using a neural network, which was trained on a digital twin without having any prior knowledge of the system dynamics, using the Proximal Policy Optimization learning algorithm. By using a multivariate Gaussian reward function for rewarding small errors between the vessel and the various setpoints, while encouraging small actuator outputs, the proposed Deep Reinforcement Learning based control scheme showed good positioning performance while being energy efficient. Both simulations and model scale sea trials were carried out to demonstrate performance compared to traditional methods, and to evaluate the ability of neural networks trained in simulation to perform on real life systems.