Neural-Based Predictive Control for Safe Autonomous Spacecraft Relative Maneuvers
Stefano Silvestrini, Michèle Lavagna
Abstract
The justification for this work and the goal of the paper is to develop an algorithm that addresses the following features: To develop a neural-based reconstruction algorithm for system dynamics identification, which allows an autonomous spacecraft to refine the on-board dynamic model as it flies, coping with unmodeled perturbations and nonlinearities. It is achieved by supervised learning of an RNN. To develop a planning algorithm that can adapt to the perturbed environments using the neural reconstructed dynamics. This prevents the failure of traditional algorithms in the presence of unmodeled terms in the dynamic model enhancing the autonomy and flexibility of the spacecraft. The task is performed using the developed MBRL method. To develop a relative trajectories prediction algorithm to ensure collision-free simultaneous reconfigurations. This is required when coordinated maneuvers are needed, where the hypothesis of the formation evolution according to natural dynamics does not hold. The neighboring trajectories are predicted by IRL and LSTM to guarantee safe reconfigurations.