AUV Position Tracking Control Using End-to-End Deep Reinforcement Learning

摘要

In this article we consider the navigation problem for an autonomous underwater vehicle (AUV) for reaching a desired way-point. The navigation problem in underwater vehicles presents major problems, the highly coupled dynamics of the vehicles and the unknown parameters of the dynamic model, make the need for complex control architectures. However, current developments in reinforcement learning show promising results for robotics applications. In particular underwater autonomous vehicles could benefit from this new techniques, achieving adaptive behavior for real-time problem solving. Based on this developments the navigation problem is solved using deep reinforcement learning, in particular the deep deterministic policy gradient. In this proposal a model free approach is used, where the raw sensor information is used as inputs to a policy network, and the outputs of this network are directly mapped to the thrusters. In addition an adaptive goal driven architecture is used to allow the agent to reach variable way points consistently. The obtained simulated results show its capacity for successfully solving AUV navigation problems.