Agile quadrotor maneuvering using tensor-decomposition-based globally optimal control and onboard visual-inertial estimation

• Main Author: Riether, Fabian
• Other Authors: Sertac Karaman.
• Format: Others
• Language: English
• Published: Massachusetts Institute of Technology 2017
• Subjects: Mechanical Engineering.
• Online Access: http://hdl.handle.net/1721.1/106777

Thesis: S.M., Massachusetts Institute of Technology, Department of Mechanical Engineering, 2016. === Cataloged from PDF version of thesis. === Includes bibliographical references (pages 123-127). === Over the last few years, quadrotors have become increasingly popular amongst researchers and hobbyist. Although tremendous progress has been made towards making drones autonomous, advanced capabilities, such as aggressive maneuvering and visual perception, are still confined to either laboratory environments with motion capture systems or drone platforms with large size, weight, and power requirements. We identify two recent developments that may help address these shortcomings. On the one hand, new embedded high-performance computers equipped with powerful Graphics Processor Units (GPUs). These computers enable real-time onboard processing of vision data. On the other hand, recently introduced compressed continuous computation techniques for stochastic optimal control allow designing feedback control systems for agile maneuvering. In this thesis, we design, implement and demonstrate a micro unmanned aerial vehicle capable of executing certain agile maneuvers using only a forward-facing camera and an inertial measurement unit. Specifically, we develop a hardware platform equipped with an Nvidia Jetson embedded super-computer for vision processing. We develop a low-latency software suite, including onboard visual marker detection, visual-inertial estimation and control algorithms. The full-state estimation is set up with respect to a visual target, such as a window. A nonlinear globally-optimal controller is designed to execute the desired flight maneuver. The resulting optimization problem is solved using tensor-train-decomposition-based compressed continuous computation techniques. The platform's capabilities and the potential of these types of controllers are demonstrated in both simulation studies and in experiments. === by Fabian Riether. === S.M.