Practical state estimation and control for autonomous quadrotor flight

Abstract

This work deals with control and autonomy for small-scale vertical takeoff and land (VTOL) unmanned aerial vehicles (UAVs), specifically with quadrotors. The contribution of this thesis is threefold. First, new robust controllers are presented with a particular emphasis on tunability in the field by operators not trained in robust control theory. One controller is a partial feedback linearizing controller, but with gains suboptimally selected to maximize a mixed H∞/H2 performance objective subject to pole placement constraints. The structure of the gains is such that they can be interpreted as PID gains and, if necessary, easily adjusted in the field based on objectives and flight conditions. The second robust controller expands on this idea to allow for PD-like tuning while maintaining robust stability and performance guarantees. It uses a standard H∞ control synthesis procedure, but an operator tuning matrix is added to the system. During controller synthesis this matrix is treated as a bounded parametric uncertainty. In the field, as long as the operator does not violate the tuning matrix bounds the original stability and performance guarantees are maintained, and the operator is able to optimize the flight performance for the specific task. The second contribution is to develop, implement, and experimentally demonstrate algorithms needed to achieve autonomous indoor flight, with an emphasis on locally stabilizing the vehicle to enable slower algorithms, such as localization, navigation, task planning, etc., to be run in the background. Where suitable, estimation and control algorithms were researched and adapted from literature, such as for the attitude observer, translation state extended Kalman filter (EKF), and vision-based control laws. In other cases new algorithms, suitable for high-speed operation on mobile computing hardware, were developed. Specifically, a new algorithm was developed within the Bayesian framework for robust velocity estimation. This uses velocity and height prior distributions, provided by the translation EKF, to estimate image-space feature location distributions, establish soft correspondences between features in the previous and current images, and finally compute the maximum a posteriori (MAP) velocity, which is used to achieve smooth control. This work is then extended to short-term, descriptor-free region tracking, which provides the rough local position information needed to maintain hover within a small area. All this is demonstrated to maintain the vehicle position without the aid of any external sensing. Finally, throughout all this work a smartphone is used as the onboard flight controller. Computation, communication, and nearly all sensing is done using the phones hardware. This demonstrates the feasibility of a smartphone to fulfill this role, and allows quadrotor systems to take advantage of the convenient packaging, relatively low cost, and frequent hardware updates provided by smartphone manufacturers.