Learning Koopman Bilinear Models using Bi-level Optimization for Quadcopter Control in Complex Dynamic Environments
Open Access
Author:
Prasetyo, Bayu
Area of Honors:
Engineering Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Junyi Geng, Thesis Supervisor Charles E Bakis, Thesis Honors Advisor
Keywords:
Koopman Quadcopter Controls Modeling
Abstract:
Quadcopters are used in an increasingly large number of applications, as they serve as a stable aerial platform for various systems. However, these applications are limited by nonlinear dynamics in certain flight conditions. The difficulty in operating in these conditions stems from how these nonlinear dynamics are typically difficult to model, which in turn makes it difficult to design a controller to operate in these conditions. However, the Koopman Bilinear Form (KBF) has risen as a model for these nonlinear conditions. The KBF model is typically learned from training data through various methods, such as the use of neural networks. However, existing Koopman neural networks typically use a Single-Level optimization (SLO) framework, which may lack accuracy. This thesis presents the use of a Koopman neural network with a Bi-Level optimization (BLO) framework to learn the dynamics of a quadcopter in ground effect flight conditions. The BLO formulation allows for a more accurate model compared to a SLO based neural network, which in turn allows for a more effective controller for operating in these nonlinear conditions. The results demonstrate that a BLO system model can be applied in a simulated control of a quadcopter flying in a ground effect environment.
