End-to-end navigation for Autonomous Underwater Vehicle with Hybrid Recurrent Neural Networks

Abstract

This paper presents a novel navigation method for Autonomous Underwater Vehicle (AUV). In respect of improving the precision of navigation, recent research mainly focused on how to reduce the system model error. However, existing optimization methods are less efficient and effective in decreasing interference of sensor deviation. Motivated by the excellent performance of deep-learning, a Hybrid Recurrent Neural Networks (Hybrid RNNs) framework is proposed to estimate the AUV position. Firstly, since the different sensors have different data frequency, this method employs unidirectional and bi-directional long short-term memory (LSTM) with multiple memory units to handle raw sensor values in a single calculation cycle. Subsequently, using the outputs of LSTMs and the time interval of the cycle above, the fully connected layers could obtain the displacements of AUV. Eventually, to verify the effectiveness of the proposed navigation algorithm, a series of evaluations have been carried out, which are based on a public dataset and real experimental data of our AUV. The evaluation results have been validated that the proposed method can reduce the interference of sensor deviation, and has better accuracy as well as fault tolerance for navigation. Meanwhile, it could also satisfy the real-time requirement.