Online Robotic Arm Control with a Deep Learning-Based EEG BCI

Abstract

EEG-based brain-computer interfaces (BCIs) have the potential to improve the lives of motor impaired patients or even healthy individuals by providing a direct line of communication with electronic devices. These systems have recently been developed to control robotic limbs using signals recorded from users' brains, offering a potential method for patient-controlled prostheses. While this advancement has the potential to improve quality-of-life for some populations, there is still much room for improvement and more work is needed to characterize the capabilities of this technology. In a recent study, we implemented a BCI system that uses deep learning (DL) to achieve high performance continuous 2-D control of a virtual cursor in online experiments. Here, we expand upon that system to control a robotic arm instead of, or in addition to, the virtual cursor. We evaluated the performance of this system using the different feedback methods over several sessions to investigate the feasibility of transitioning from a virtual application to a physical device. Our results show that subjects were able to effectively control the robotic arm and achieved similar performance compared to the virtual cursor control tasks after training. These results provide evidence that BCI systems can transition from virtual tasks to controlling physical devices without a substantial loss in performance, and provide a performance benchmark for continuous BCI robotic arm control using online DL-based decoding.