Data-Driven Modeling and Control of Robotic Systems

Using data-augmented modeling for prediction and control

Project Description

In robotics, first-principle models often describe the system at hand only with limited accuracy. Sources of mismatches with respect to their nominal behavior can be hard-to-model physical phenomena, such as contact forces and friction, or by the use of lower-precision components commonly found in contemporary, cost-effective manufactured robots. These limitations motivate the incorporation of data-driven modeling approaches to better capture these systematic uncertainties. By leveraging data-inferred models, it is possible to improve predictive accuracy and to integrate these models into optimization-based controllers, thereby enhancing control precision.

To this end, we explore different data-driven modeling approaches, specifically investigating Gaussian process regression or Extended Dynamic Mode Decomposition on a Koopman background, to develop data-based model extensions of wheeled mobile robots using real-world data. One of our main objectives is it to integrate the obtained data-inferred models within predictive control settings, where the controllers’ practical applicability is examined in the institute’s laboratory with real-world robots. Special focus is also placed on the combination of data-based modeling and the specific challenges of controlling nonholonomic vehicles.

00:20
To the top of the page