Learning Collision Avoidance of Ship Manoeuvring based on Gated Recurrent Unit

Abstract

Unmanned surface vehicle (USV) has progressed quickly in recent decades, with widespread research and practical applications in academic and industry circles. Collision avoidance is a fundamental capability of USVs. It is extremely challenging to develop an ideal and sophisticated collision avoidance algorithm for USV in complex environments and practical offshore situations. However, a supervised learning method provides the USV a way of learning the process of avoidance and imitating a human pilot maneuvering to navigate in the real environment. This study analyzes the relative relationships and features of the own and target ships in the avoidance process firstly. And then a Gated Recurrent Unit (GRU) recurrent neural network model is constructed. Maneuvering commands during avoidance by human pilots are utilized as tags for training. Finally, the validity of method is proven by performing navigation experiments. In particularly, we also compare the effectiveness of GRU with Long Short-Term Memory (LSTM) network which is also a kind of recurrent neural network. The experimental results indicate that the proposed GRU model is better than the LSTM model, and the USV can autonomously navigate during the collision avoidance process by using the well-trained GRU model, reaching a level similar to that of human pilots.