Spacecraft Formation Relative Trajectories Identification for Collision-Free Maneuvers using Neural-Reconstructed Dynamics

Abstract

A Model-Based Reinforcement Learning algorithm (MBRL) is proposed to perform nearly-optimal reconfiguration and maintenance in distributed formation flying spacecraft. In addition, two algorithms, namely Inverse Reinforcement Learning (IRL) and Long Short-Term Memory (LSTM) network, are proposed to reconstruct and predict neighbouring future trajectories to perform collision-free maneuvers when simultaneous reconfiguration occurs, i.e. all spacecrafts evolve following non-natural trajectories. The dynamics reconstruction, used for planning exploits a Recurrent Neural Network, whereas the IRL uses a nested short-horizon optimization to approximate the cost function. The paper compares the proposed algorithm with a baseline Model-Predictive Control approach. In particular two mission concepts are studied as application scenarios.