Perching with fixed wings

Abstract

Human pilots have the extraordinary ability to remotely maneuver small Unmanned Aerial Vehicles (UAVs) far outside the flight envelope of conventional autopilots. Given the tremendous thrust-to-weight ratio available on these small machines [1, 2], linear control approaches have recently produced impressive demonstrations that come close to matching this agility for a certain class of aerobatic maneuvers where the rotor or propeller forces dominate the dynamics of the aircraft [3, 4, 5]. However, as our flying machines scale down to smaller sizes (e.g. Micro Aerial Vehicles) operating at low Reynold's numbers, viscous forces dominate propeller thrust [6, 7, 8], causing classical control (and design) techniques to fail. These new technologies will require a different approach to control, where the control system will need to reason about the long term and time dependent effects of the unsteady fluid dynamics on the response of the vehicle. Perching is representative of a large class of control problems for aerobatics that requires and agile and robust control system with the capability of planning well into the future. Our experimental paradigm along with the simplicity of the problem structure has allowed us to study the problem at the most fundamental level. This thesis presents methods and results for identifying an aerodynamic model of a small glider at very high angles-of-attack using tools from supervised machine learning and system identification. Our model then serves as a benchmark platform for studying control of perching using an optimal control framework, namely reinforcement learning. Our results indicate that a compact parameterization of the control is sufficient to successfully execute the task in simulation.