Logo image
Back
Developing an optical brain-computer interface for robot control
Dissertation   Open access

Developing an optical brain-computer interface for robot control

Alyssa Marie Batula
Doctor of Philosophy (Ph.D.), Drexel University
Apr 2017
DOI:
https://doi.org/10.17918/etd-7541
pdf
Batula_Alyssa_20179.97 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Electrical engineering Brain-computer interfaces Near infrared spectroscopy Machine Learning
The ability to direct a robot using only human thoughts could provide a powerful mechanism for human-robot interaction with a wide range of potential applications including medical robotics, search-and-rescue operations, and industrial manufacturing. Brain-computer interfaces (BCIs) are systems that allow the user to control a computer with only their thoughts, providing a promising research area for new methods of robotic control. They could be used to control the navigation of a robotic wheelchair, an assistive or telepresence robot that performs errands, or even the movement of a prosthetic limb. In this work I present the design and evaluation of the first BCI to use four imagined movements recorded via functional near-infrared spectroscopy (fNIRS) to control both a virtual and a physical robot. The BCI is used to navigate the robot to a goal location in a room, a prototype and initial step towards remote control of a telepresence or assistive robot. Four imagined movement tasks (tapping of the left hand, right hand, left foot, and right foot) are mapped to high-level commands (turn left, turn right, walk forwards, walk backwards) to direct the robot. The ability to reliably distinguish multiple mental tasks is essential for use in a practical BCI. In an offline analysis I compare the activation patterns generated during both motor imagery and motor execution (actual movement). This is the first analysis of the activation patterns recorded via fNIRS separately for left and right foot motor imagery tasks. Signal processing, feature extraction, and machine learning methods are integral parts of BCI design. In an additional offline analysis I compare classification results using eight methods of signal preprocessing that have been suggested for use in fNIRS BCIs. I also provide comparisons of two commonly-used classifiers in BCIs as well as feed-forward and convolutional neural networks. Additionally I present the results of a five-class classification task, adding a resting state to the four motor imagery tasks, which could potentially increase the number of inputs available to the BCI.

Metrics

66 File views/ downloads
115 Record Views

Details

Logo image