Predictive Control of Constrained Nonlinear Systems

Abstract

Because of advanced autonomous capabilities, mobile robot systems are transitioning from usage in controlled laboratory and factory settings to deployment in unstructured and increasingly complex environments. However, robustness of these field robots to environmental unknowns at the time of deployment remains a challenge. For example, in aerial infrastructure inspection, control of micro air vehicles is challenging due to induced transient flow conditions and varied lighting conditions that degrade state estimation. To safely navigate within these settings, an efficient and adaptive control strategy is essential to mitigate risk and increase system robustness.

This thesis work examines a computationally efficient strategy for solving nonlinear model predictive control problems by learning from past experience. Efficiency is achieved by constructing a library of controllers that map the state, reference, and dynamics model to the locally optimal feedback control laws. An affine dynamics model is combined with an online learned component that estimates nonlinearities and perturbations to accurately model the system. The safety, accuracy, and robustness of this control technique is evaluated by comparing the tracking performance against classic reactive control strategies.