We explore the use of machine learning to reproduce human behaviors as a means to facilitate more natural interaction with autonomous systems. In this paper, we use simplified formation flight as a domain to explore whether a neural net can effectively be used to learn a behavior from human pilot example data. We train the neural net from a virtual reality-based flight simulator, evaluate its functionality in high-fidelity software and hardware in the loop simulation, and demonstrate successful operation with flight testing. This work validates that our approach successfully transfers from simulation to real-world conditions, generalizes well to situations it was not explicitly trained for, and responds well to noise and perturbations. Our results show the surprisingly small neural net required for this problem and its corresponding quick execution time on commodity cell-phone class hardware, as well as estimates of the quantity of training data required….Read more

Posted by Rockwell Collins