Quantifying Impact of Information Sharing on Satellite Navigation

Communication between satellites is an important component of collision-free navigation operations. The work in this paper is motivated by the premise that each satellite has more reliable knowledge of its own state compared to its direct observations of others. The authors aim to leverage this shared, enhanced information among satellites to enable autonomous coordination and collision avoidance. They rely on multi-agent reinforcement learning (MARL) to model multiple autonomous satellites interacting in a shared environment with limited observation and communication ranges. First they demonstrate the effectiveness of their method against comparable MARL algorithms. To address the computational expense of MARL, they investigate the effectiveness of transferring a model trained on a ground environment to one trained in space. Despite the fact the two environments have very different dynamics, they show that the transfer learning model outperforms a model that is trained directly on the space environment. Furthermore, even in the presence of additional perturbations, the authors’ transfer learning model is able to accelerate training and achieve comparable rewards. With the effectiveness of their method explored, they then evaluate the benefits of information sharing by studying collision rates and path planning. Their results indicate that sharing maneuver information leads to lower collision rates and more efficient path planning.