Applying Deep Neural Networks to Improve UAV Navigation in Satellite-less Environments

Abstract

In this work, a novel algorithm based on machine learning to tackle the navigation problem of Unmanned Aerial Vehicle (UAV) in satellite-less surroundings is presented. The proposed network is trained by exploiting the outputs of a Generalized Trust Region Sub-problem (GTRS) method, after which a deep Long Short-Term Memory (LSTM) model is applied to predict the drone’s position. We present here our preliminary findings which indicate high potential and utility of machine learning tools for the problem of interest. More precisely, the results reveal improved accuracy, while matching the execution time of the GTRS method. Moreover, the proposed method also reveals lower susceptibility to noise; in practice, these two factors combined can be the difference between the UAV colliding with an obstacle or not.