Risk-based Convolutional Perception Models for Collision Avoidance in Autonomous Marine Surface Vessels using Deep Reinforcement Learning

Abstract

In this work, we propose a novel policy network architecture for model-free Reinforcement Learning (RL)-based path-following and collision avoidance in marine surface vessels. By applying convolutional neural networks (CNNs) for mapping LiDAR-like distance measurements to Collision Risk Indices (CRIs), we evaluate the utility of risk-based pretraining of CNN feature extractors prior to RL. Where previous works required hand-crafted preprocessing of high-resolution distance measurements to train an autonomous RL agent successfully, the proposed approach achieves this goal in a data-driven fashion. Ultimately, we propose future directions to improve CNN-based perception models for collision avoidance in range sensing applications.