Optimizing Parameters for Hybrid DNN-UKF State Estimation in Autonomous Air Refueling

The future of the United States Navy's (USN) carrier airwing hinges on effective Uncrewed Aerial Vehicles (UAVs), whose autonomy critically depends on robust aerial refueling systems. This work seeks to refine the process and measurement noise parameters in a hybrid Deep Neural Network (DNN) and Kalman Filter (KF) framework to improve drogue tracking reliability under challenging operational conditions. The proposed study employs a structured experimentation protocol combining simulated lab environments-featuring robotic arms replicating drogue motions-along with video datasets from actual refueling operations. Preliminary results obtained via a trial-and-error tuning approach indicate promising performance, achieving an overall RMSE of 0.155 m. Building on these encouraging findings, we seek to implement systematic fine tuning methods, specifically grid search and Bayesian optimization, to further reduce RMSE and enhance system accuracy and robustness. The ultimate goal is to advance the operational readiness of autonomous aerial refueling, laying the groundwork for the next generation of USN carrier-based aviation.