Learning Quadrotor Maneuvers from Optimal Control

Abstract

In this diploma thesis, a novel approach for quadrotor maneuver generation is presented. Using Dynamic Movement Primitives, offline computed solutions of optimal control problems are learned and generalized. A mesh of solutions is created, learned and the weights of the Gaussian basis are interpolated linearly. To generalize trajectories with different durations, an enhanced temporal scaling method is applied. The developed framework was implemented using a planar quadrotor model, but can be extended directly to a more detailed model. It allows to generate different optimal maneuvers, for instance point-to-point with arbitrary goals or flips with defined heights, and in-flight modifications, e.g. to avoid obstacles. States of the quadrotor can be defined at instants during the maneuver while in general no assumptions about the trajectory profiles are made. The trajectories implicitly fulfill given input and state constraints and different optimality criteria can be taken into account. To evaluate the developed method, simulations and experiments with a real quadrotor in an indoor test environment were performed.