Comparing DNN Performance to Justify Using Transference of Training for the Autonomous Aerial Refueling Task
Dillon Miller, Violet Mwaffo, Donald H. Costello
Abstract
In an effort to modernize the fleet, the United States Navy is looking to significantly increase the number of unmanned aircraft deployed within a carrier air wing. Yet, no method to certify the autonomous refueling of uncrewed aerial platforms has been publicly released. Ongoing research efforts at the United States Naval Academy (USNA) are investigating certification evidence that will allow a deep neural network (DNN) to enable the autonomous aerial refueling task. This poster paper highlights an investigation into developmental flight test videos of an aircraft refueling from a KC-130 tanker and from a tanker configured F/A-18 jet. In this paper, we evaluate a KC-130 trained DNN and a F/A-18 trained DNN against a F/A-18 data set that was not used in training either DNN. This procedure was aimed at determining whether the resources required to gather training data on each tanker aircraft taken separately are justified or if the performance of the DNN trained on a similar aircraft dataset is sufficient for the task.