Autonomous operation of multiple unmanned aerial vehicles (UAVs) becomes difficult when communication is intermittent and global navigation satellite system (GNSS/GPS) measurements are degraded. This study develops and evaluates a decentralized three-layer control architecture in which stabilization, onboard computation, and inter-UAV networking are separated while each UAV retains local perception and decision-making capability. The implementation combines a Pixhawk/Matek flight controller with a Raspberry Pi 4 or NVIDIA Jetson Nano companion computer, visual and LiDAR sensing, visual-inertial state estimation, ROS 2 middleware, and a peer-to-peer Wi-Fi mesh. Swarm coordination uses local neighbor information, a modified Boids model, predictive collision avoidance, and barrier-function-based safety constraints. The system was evaluated in Gazebo, AirSim, and PX4 software-in-the-loop environments and in controlled field experiments. The reported measurements show mesh-network delays of approximately 30–80 ms, localization RMSE of 0.32–0.48 m under degraded GPS, and total obstacle-response times below 1 s, with direct field measurements of approximately 0.5–0.6 s. Across the reported field campaign, 27 of 30 missions were completed without operator intervention. The results indicate that distributing perception, state estimation, and local control across UAVs can maintain useful swarm coordination despite temporary communication and positioning degradation. The main limitation is the small experimental scale and the use of a homogeneous swarm; therefore, the results should be interpreted as validation of the architecture rather than evidence of large-scale swarm scalability.