A Hybrid LSTM-based Neural Network for Satellite-less UAV Navigation

This work proposes a new algorithm to address the problem of unmanned aerial vehicle (UAV) navigation in satellite-less environments by combining machine learning with existent model-based methods. The proposed network model is trained by using the predictions of two estimators, one based on a Generalized trust region sub-problem (GTRS) framework and the other one founded on a Weighted Least Squares (WLS) principle. The solutions of these two estimators are then fed to two Long Short-Term Memories (LSTMs) to create models whose predictions are averaged to achieve the final prediction output. Our numerical results show favorable performance of the new network, obtaining improved accuracy and higher robustness to noise when compared with the individual counterparts of the network used in the training phase. Consequently, the proposed method offers safer an more reliable navigation of the UAV in satellite-less environments.