From Demonstration to Flight: Realization of Autonomous Aerobatic Maneuvers for Fast, Miniature Fixed-Wing UAVs

Performing agile maneuvers autonomously remains a challenge for miniature fixed-wing unmanned aerial vehicles (UAVs) because they are required to fly along complex trajectories while approaching their physical limits. This article proposes an integrated maneuver planning and tracking method to address this issue. The developed approach is built upon the learning from demonstration (LFD) framework. First, an experienced pilot performs several sophisticated maneuvers manually. Then, instead of directly controlling the UAV to track the demonstrated trajectory, we split the teaching maneuvers into multiple segments. A novel dynamic motion primitive in terms of a unit dual quaternion (DQ-DMP) is then developed to encode and store those segments' rotational and translational features. Besides, a robust connecting method for the learned DQ-DMPs is introduced, resulting in a smooth maneuvering trajectory that even the pilot has not taught yet. Additionally, a cascaded dynamic inversion control scheme with stable tracking performance is also developed. Our system is integrated into an autonomous fixed-wing drone and implements severalcomplex maneuvers in the outdoor environment. The results of comparative experiments and outdoor flight reveal that our approach can learn, adapt and generalize maneuvers in real-world scenarios.