Data-Driven Deformable Object Manipulation Using Koopman-based Learning Approach
A universal data-driven learning and control framework
Background: Koopman operator is an infinite-dimensional linear operator that can govern the evolution/transition of observables (lifting nonlinear dynamics into a higher dimensional space). It provides the capability to represent a highly nonlinear dynamic system using an infinite-dimensional linear operator to achieve global linearization of nonlinear dynamic systems. It is completely data-driven (i.e., does not require the knowledge of the underlying dynamics).
Motivation and Goal: In this research, we are aiming at addressing the challenges associated with deformable object manipulation originating from uncertain deformation behavior due to highly nonlinear dynamics, uniform or nonuniform physical properties - including geometry, stiffness, mass distribution, viscoelasticity, friction, ect. The goal of this project is to develop a data-driven control framework combining Koopman-based learning and model predictive control to manipulate a deformable object with unknown physical properties to autonomously accomplish a predefined task in an environment with potential disturbances.
Koopman-based Learning Control Framework:
Test the Generality of the Algorithm Using a Differential Drive Robot:
