An experimental comparison of Deep Learning strategies for AUV navigation in DVL-denied environments

Accurate and robust navigation and localisation systems are critical for Autonomous Underwater Vehicles (AUVs) in order to perform missions in challenging environments. However, since the Global Positioning System (GPS) is not available in the underwater domain, the localisation task is commonly fulfilled by integrating direct linear speed readings provided by a Doppler Velocity Log (DVL) over time. As a consequence, DVL failures or fallacies and DVL-denied environments may arise as unexpected causes for severe malfunctions of the whole navigation system. Motivated by these considerations and the outstanding performance of Deep Neural Networks (DNNs) in supervised regression problems, a Deep Learning (DL)-based approach has been developed to estimate the vehicle's body-frame velocity, without canonically employing DVL measurements, in a Dead-Reckoning (DR) navigation strategy. In particular, this work will describe the whole framework, starting from the data gathered by the AUVs of the National Oceanography Centre (NOC) during different field campaigns, through to the data pre-processing and the inference of the predicted velocity. Finally, a comprehensive offline comparison between different DL-based models is presented to assess the validity of the proposed approach.