Real-time optimal control via Deep Neural Networks: study on landing\n problems

摘要

Recent research on deep learning, a set of machine learning techniques able\nto learn deep architectures, has shown how robotic perception and action\ngreatly benefits from these techniques. In terms of spacecraft navigation and\ncontrol system, this suggests that deep architectures may be considered now to\ndrive all or part of the on-board decision making system. In this paper this\nclaim is investigated in more detail training deep artificial neural networks\nto represent the optimal control action during a pinpoint landing, assuming\nperfect state information. It is found to be possible to train deep networks\nfor this purpose and that the resulting landings, driven by the trained\nnetworks, are close to simulated optimal ones. These results allow for the\ndesign of an on-board real time optimal control system able to cope with large\nsets of possible initial states while still producing an optimal response.\n