Assessment of BCI Performance for Human-Robot Interaction

Abstract

Brain-computer interface (BCI) devices can enable individuals suffering from motor impairments to live a more comfortable life. By monitoring brain activity, BCIs translate neural processes to control signals for external devices such as robot arms. One method is by detecting and classifying electroencephalography (EEG) signals related to motor effort or execution while minimizing errors due to artifacts. This paper proposes a comprehensive framework for accurate binary motor execution classification between relaxed and activated curl states of the arm using raw EEG signals, using filter bank common spatial pattern (FB-CSP) feature extraction, and machine learning techniques. We tested our approach on 10 healthy subjects aged 22-33 to compare CSP and FB-CSP feature extraction techniques alongside machine learning (ML) techniques of support vector machine (SVM), logistic regression (LR), linear discriminant analysis (LDA) and Naïve Bayes (NB) to identify the best performer. Two types of tests were conducted, including offline and live testing. Human-Machine Interaction (HRI) was accomplished by mapping decision thresholding of the ML probability score outputs to produce a mirrored pose of a robot arm. For our experiments, SVM achieved the highest validation accuracy with an average of 89.09%. Additionally, two types of tests were conducted, including offline and live testing. Furthermore, our results suggest that providing visual feedback is important for motivating and refining motor intent signals for more consistent BCI control. These initial findings highlight the promise of non-invasive BCIs in HRI applications, with ongoing research aimed at further improving accuracy, functionality, and usability to optimally meet the needs of targeted assistive technologies.