Controller design for quadrotor UAVs using reinforcement learning

Quadrotor UAVs are one of the most preferred type of small unmanned aerial vehicles because of the very simple mechanical construction and propulsion principle. However, the nonlinear dynamic behavior requires a rather advanced stabilizing control of these vehicles. One possible approach that relaxes the difficult task of nonlinear control design is the application of a learning algorithm that allows the training of suitable control actions. Here we apply reinforcement learning as one form of unsupervised learning. In this paper, we first propose a nonlinear autopilot for quadrotor UAVs based on feedback linearization. This controller is then compared to an autopilot which has been learned by reinforcement learning using fitted value iteration with regard to design effort and performance. First simulation and experimental results underline the outcome of this comparison.