Deep Learning-Based Estimation of UAV-Swarm Communication Constraints

Abstract

UAV swarms must often follow predefined trajectories while preserving relative positions under changing communication conditions. We estimate communication failure rate and maximum range from a compact set of high-level swarm metrics. The proposed approach uses a fully connected network driven by temporal summaries shape_error, leader_off, presence_error, size_formation, and comm_delay. It predicts control-relevant communication parameters that are otherwise unobservable, using only a minimal set of conventional, objectively measurable variables. Simulations with Raft-based decentralized control in GrADyS-SIM NextGen across diverse conditions show strong predictive accuracy, with R₂>0.83 and low test loss for both targets. The method is non-intrusive and suitable for online monitoring of communication health. Contributions include a compact metric set, independent estimation of failure and range, and validation across varied scenarios.