Another work that I'm dealing with is Learning of Acrobatic Maneuvers from Pilot Demonstrations. In this work, we have an expert pilot who performs different acrobatic manuvers. We use these data to infer the essential parts of the shown maneuver. Then, based on this learnt model, the robot is able to execute the manuver with the same performance (or even higher) than the pilot. To our knowledge, this is one of the fastest ways of implementing any complex maneuver on UAVs.
Figure Black line: trajectory extracted from the demonstration (blue points in cross); Green area: a region of 3 times the variance around the extracted trajectory. Left: 3D visualization of the generated trajectory (the starting point of all trajectories is shifted into origin). Right: aircraft velocity along time.