Zero-Shot Benchmarking of Monocular Depth Estimation Networks for Autonomous Aerial Refueling

Autonomous aerial refueling is essential for extending the operational reach of Unmanned Aerial Systems (UASs) in contested environments, and reliable monocular depth estimation is a key enabler for GPS-denied terminal-phase guidance. This study benchmarks three state-of-the-art monocular deep neural networks (MiDaS, AdaBins, and NeWCRFs) for estimating drogue–probe separation distance using imagery collected in a controlled laboratory environment. All models are evaluated in a zero-shot setting, without any fine-tuning, to quantify their intrinsic generalization capability and to assess the impact of domain mismatch on depth-prediction fidelity. Performance is measured using absolute and normalized depth error, overall RMSE, and inference latency. Results show that NeWCRFs trained on the NYU dataset provides the highest accuracy and stability across all stand-off distances, while AdaBins offers fast inference at the expense of significantly larger errors. The findings reveal a clear accuracy–latency trade-off that currently precludes real-time deployment. These results inform the design of lightweight, scalable perception modules for future autonomous aerial refueling systems and motivate the development of domain-matched training data and latency-optimized depth-estimation architectures.