Transfer Learning for Space Traffic Management

We explore space traffic management as an application of collision-free navigation in multi-agent systems where vehicles have limited observation and communication ranges. We demonstrate using simulations that GNNs are both transferable (i.e., they can use initialized weights from analogous ground-based environments), and scalable with respect to the number of agents. We find that transfer learning can improve sample efficiency and performance compared to both the case where the model is trained directly on a space-based environment, as well as other baseline MARL approaches. Finally, we use our model to quantify the value of sharing maneuver information between satellite operators in order to improve decision-making.