Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.


Online Model Learning in Autonomous Racing Using Gaussian Processes


S. Ercan

Semester Thesis, SS17

In this thesis we investigate how Gaussian Processes can be employed in online model learning in autonomous racing. Success of high-speed, real-time autonomous racing depends on the availability of a good car model. However, the model dynamics change throughout operation due to factors such as wearing of the tires, aging of the motor, dust on the track or batteries with different state of charge. These model variations can significantly degrade the accuracy of controller predictions and in turn lead to poor driving performance. In this presentation, we show how to use Gaussian Processes to learn variations of the underlying dynamical model in a black-box approach, that is, purely from recorded data. We then demonstrate that the resulting GP model compares favorably to the existing grey-box bicycle model in open loop trajectory predictions. Furthermore, we integrate the GP model into the existing Model Predictive Contouring Controller framework and evaluate the resulting controller in simulations. The controller is close to real-time feasible and we also provide an initial exploration of efficient training data selection in order to retrieve real-time feasibility.

Supervisors: Alex Liniger, Marius Schmitt, John Lygeros


Type of Publication:

(13)Semester/Bachelor Thesis

File Download:

Request a copy of this publication.
(Uses JavaScript)
% Autogenerated BibTeX entry
@PhdThesis { Xxx:2017:IFA_5676
Permanent link