Performance of a Deep Neural Network Trained on Professionally Generated Gaussian Splat Imagery for Autonomous Aerial Refueling

The integration of uncrewed aerial systems (UAS) into carrier air wings by the United States Navy requires that these systems possess the capability to perform autonomous aerial refueling (AAR) in GPS/RF-denied environments. Deep neural networks (DNNs) trained using flight test imagery have demonstrated promising results for drogue detection and tracking. However, collecting and labeling flight test data is costly and logistically challenging. This work investigates whether artificial imagery generated from gaussian splatting techniques can serve as an effective alternative to flight test data. Using the University of Maryland’s Holocamera Studio, images of an F/A-18 aerial refueling drogue were processed using background subtraction, neural modeling, and gaussian splatting to produce synthetic videos. Labeled images extracted from these videos were augmented and used to train a YOLOv8n model under the same conditions as a second model trained on real flight-test footage. Both models were evaluated on an independent set of 102 real flight-test images. Results show that the DNN trained on artificial data achieved an overall mAP50–95 of 0.847, compared to 0.889 for the DNN trained on real flight test data, indicating comparably strong performance. Notably, the artificial-data DNN outperformed the real-data model in coupler detection (0.819 vs. 0.804 mAP50–95). These findings demonstrate that professionally generated datasets can provide a low-cost, scalable alternative for training DNNs in safety-critical aerial refueling tasks, reducing the dependence on expensive and logistically complex flight-test data collection.