Date of Award

2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical Engineering

Committee Chair

Farbod Fahimi

Committee Member

Avimanyu Sahoo

Committee Member

Rohan Sood

Committee Member

Howard Chen

Committee Member

Chang-kwon Kang

Research Advisor

Farbod Fahimi

Subject(s)

Adaptive control systems, Self-organizing systems, Neural networks (Computer science), Machine learning

Abstract

Aerospace and robotic systems perform various tasks in uncertain, dynamic environments. Further, the data that a system encounters in the real-world is often the most valuable. The most advanced aerospace and robotic systems of the future will be able to learn online, during operation, from this real-world data. Aerospace and robotic systems are often expensive and difficult to model, and the controllers of these systems thus require rigorous control proofs and safety guarantees. Recently, general research in artificial intelligence and machine learning (AI/ML) has made significant strides in developing learning-based systems and controllers. However, much of this research has focused on optimizing control performance, robustness, or prediction accuracy, without considering the stability and safety requirements for control of real-world aerospace and robotic systems. Additionally, the optimization objectives of adaptive control and machine learning, adapting parameters over time to achieve some desired goal or performance, are often closely related. Thus, the theory and mathematical rigor of adaptive control can be used to augment popular AI/ML tools for stability guarantees and online learning. This dissertation discusses research progress in utilizing AI/ML tools, namely deep neural networks, together with adaptive control to achieve provably-stable online learning and optimization. Part I first describes “learning for control,” where neural networks are stably used in a fully online model-based nonlinear controller. The derived controller is shown to desirably control robotic arms, spacecraft, and quadcopters under various disturbances and model uncertainties with limited a priori modeling. Next, Part II describes “control for learning,” where control-theoretic techniques are used to stably update deep neural network parameters online. The proposed update law is shown to give desirable performance when deep neural network outputs are used in predicting or controlling dynamical systems, especially under domain shift from the training distribution to the target distribution, common in forecasting and sim-to-real transfer of control policies. Throughout this dissertation, the connections between machine learning and adaptive control are explored, with each field acutely poised to benefit the other.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.