Learning Quadrotor Maneuvers From Optimal Control and Generalizing in Real-Time

Abstract

In this paper, we present a method for learning and online generalization of maneuvers for quadrotor-type vehicles. The maneuvers are formulated as optimal control problems, which are solved using a general purpose optimal control solver. The solutions are then encoded and generalized with Dynamic Movement Primitives (DMPs). This allows for real-time generalization to new goals and in-flight modifications. An effective method for joining the generalized trajectories is implemented. We present the necessary theoretical background and error analysis of the generalization. The effectiveness of the proposed method is showcased using planar point-to-point and perching maneuvers in simulation and experiment.