Lupashin, S.D'Andrea, R.2012-08-232012-08-232012-08-23201110.3182/20110828-6-IT-1002.03205https://infoscience.epfl.ch/handle/20.500.14299/84990We describe a process for enabling quadrocopters to perform and improve upon aerobatic maneuvers. We describe such maneuvers as a set of desired keyframes and a parametrized input trajectory. The full state trajectory of the vehicle is left unspecified - only predefined partial-state keyframes are used to measure errors and to refine the primitive. A first-principles model is used to find nominal trajectory parameter values and a first-order correction matrix. We apply this method to extending previous work on vertical-plane 2D adaptive flips to a fully 3D adaptive maneuver. We also show how this method can be applied to finding trajectories for flips with matching non-zero initial and final velocities. Preliminary results are presented from simulation and from quadrocopters in the ETH Flying Machine Arena.Flying robotsAdaptive controlModel approximationLearning control.text::conference output::conference proceedings::conference paper