Federated multi-agent reinforcement learning for interference-aware precoding in LEO integrated communication and navigation

Low-earth-orbit (LEO) constellations are expected to deliver global broadband while broadcasting positioning beacons, enabling integrated-communication-and-navigation (ICAN) services. A central difficulty is to shape hundreds of multi-satellite beams so that aggregate throughput is maximised without violating strict ranging-signal requirements, all under limited feeder-link capacity and privacy constraints. This paper develops a federated multi-agent reinforcement-learning (FedMARL) framework for distributed downlink precoding in large LEO constellations. Each satellite runs a lightweight actor–critic pair that updates its precoder from local observations, while sparsified and differentially-private model increments are aggregated on the ground. The design enforces hard per-satellite power limits through an in-network projection layer and jointly handles data and navigation streams. Simulation results demonstrate that FedMARL approaches centralised throughput, substantially lowers navigation outage compared with per-satellite zero-forcing, and retains high positioning success even under aggressive gradient compression and privacy budgets. The proposed approach thus offers a scalable, power-feasible and privacy-preserving solution for next-generation ICAN mega-constellations.