FedS-SLAM: A Federated Semantic Collaborative SLAM System for UAV Swarms in Dynamic

Abstract Existing SLAM technologies struggle to meet the demands of drone swarms in dynamic, GPS-denied environments. To address this, we propose FedS-SLAM, a federated semantic cooperative SLAM system designed specifically for drone swarms. It innovatively employs a federated learning architecture to drastically reduce communication overhead by exchanging only lightweight "federated features." The system's core is a deeply coupled optimization loop where the SLAM back-end provides high-quality data to train a shared model via federated learning, which in turn enhances the SLAM front-end. This synergy enables the precise filtering of dynamic objects, significantly improving localization robustness. Experiments on multiple public benchmark datasets prove that FedS-SLAM outperforms state-of-the-art systems in both trajectory accuracy and map quality.