high performance computing on graphics processing units: hgpu.org

Over the past few years, in the huge field of Artificial Intelligence (AI), new Machine Learning techniques are playing a central role, proving to be very powerful and versatile. For this reason, it is expected that they could become protagonist of space applications and they are already under study. Thanks to the large availability of data, neural networks are able to elaborate and extract useful information to predict from sequences of data or to classify new ones. The goal of this work is to demonstrate how it is possible to lighten space navigation, replacing computational heavy algorithms with a well trained neural network. This proves to be a very useful tool in proximity operations with celestial objects. More in detail, this research is focused on the development of neural networks able to perform a lunar landing, exploiting an optic navigation, i.e. Moon surface images. Recurrent Neural Networks (RNN) have been developed, in which the descent trajectory state is the input and the relative control action is the output. Considering the optimal achieved results, the analysis has been carried forward with the study of Convolutional Neural Networks (CNN), where the lunar surface images are the inputs and the relative control actions are the outputs. Two kinds of landings have been considered: the first one is a pure vertical (1D) landing, faced exploting only the images; the second one is a planar (2D) landing. In this last case, both neural networks have been used (RNN and CNN). The results are particularly satisfying even if the limits of the adopted approach emerged. In fact, in the final part of this thesis new solutions for future developments are proposed.