Federated Learning on Edge Devices in a Lunar Analogue Environment

Future lunar space missions will generate vast amounts of data via multiple edge devices, which can be used to train machine learning models for exploration and navigation purposes. Given the bandwidth limitations of space assets, federated learning can enable the training of machine learning models on multiple edge devices simultaneously, with greater data efficiency than if the training data were transferred to ground stations on Earth. To explore the use of federated learning in a lunar operations scenario, we implemented federated learning in a lunar analogue environment across a network of distributed lunar nodes, including rovers, and trained a flight-ready deep learning lunar terrain classifier, MoonNet Federated. We offer our insights as to how to construct an optimal federation for a particular concept of operations, and highlight challenges and future directions for the use of federated learning in lunar space missions.