Federated multi-agent mapping for planetary exploration

A growing challenge in multi-agent robotic exploration lies in managing vast quantities of decentralized, heterogeneous streaming data produced during interactions with dynamic environments. Simultaneously, agents should acquire versatile skill sets through diversified experiences in varying contexts. Centralizing or transmitting entire largescale data poses impracticalities regarding storage capacity and communication bandwidth. Therefore, we investigate federated learning techniques applied to distributed mapping as a promising strategy to enable multi-agent learning without requiring transmission or comprehensive datasets. We define an approach based on implicit neural mapping, meta-initialized on Earth mapping data and show a capacity for models to generalize to novel domains such as the Martian surface or glaciers. Furthermore, we evaluate the effectiveness of this distributed mapping approach in relevant real-world deployment scenarios.