An Image-based Brain-Computer Interface Using the P3 Response

Abstract

We present a new image-based interface for using electroencephalographic signals (EEG) to control a humanoid robot. In order to make maximal use of a low-bandwidth brain computer interface (BCI), we use a humanoid robot that has sophisticated capabilities such as navigating to different locations in its environment and interacting with objects. The robot communicates with the BCI by sending images of discovered locations and objects as candidates for interaction. The BCI uses these candidates and exploits the P300 response for detecting the user's selection. We describe the design of the interface, and present results from a 9-user study that characterizes the performance, generalization and training needs of the BCI. Our results indicate that a 4-class selection can be performed in 5 seconds with as low as 5% error. Further, a short training period of 3-4min provides near-optimal accuracy.