Automatic ship collision avoidance using deep reinforcement learning with LSTM in continuous action spaces
Ryohei Sawada, Keiji Sato, Takahiro Majima
Abstract
Abstract This paper presents an automatic collision avoidance algorithm for ships using a deep reinforcement learning (DRL) in continuous action spaces. Obstacle zone by target (OZT) is used to compute an area where a collision will happen in the future based on dynamic information of ships. Agents of DRL detects the approach of multiple ships using a virtual sensor called the grid sensor. Agents learned collision avoidance maneuvering through Imazu problem, which is a scenario set of ship encounter situations. In this study, we propose a new approach for collision avoidance with a longer safe passing distance using DRL. We develop a novel method named inside OZT that expands OZT to improve the consistency of learning. We redesign the network using the long short-term memory (LSTM) cell and carried out training in continuous action spaces to train a model with longer safe distance than the previous study. The bow cross range in collision detection proposed in this paper is effective to COLREGs-compliant collision avoidance. The trained model has passed all scenarios of Imazu problem. The model is also validated by a test scenario which includes more ships than each scenario of Imazu problem.