Advancing Brain-Computer Interfaces for Robotic Embodiment Systems

Thanks to the rapid development of brain-computer interfaces (BCIs), users can control and communicate with external devices without performing a single muscle movement. BCIs allow this by detecting the control and communication signals directly at their source, i.e. the brain. This renders BCIs highly applicable in different domains and of special importance to users with physical impairments. However, despite the great progress made to date, one should not expect BCIs to replace keyboards in the near future due to several factors. Firstly, the process of decoding brain activity remains error-prone due to the presence of noise and nonstationarity in the measured brain activity. Secondly, the bandwidth and bit rate of current BCIs compare poorly to conventional user interfaces. Furthermore, immersive robotic embodiment applications, which are the driving force behind this work, pose additional challenges to BCIs due to the varying time scales in which robots operate. To circumvent the current limitations, efforts were made firstly to improve the accuracy and reliability of the detection of the brain signals used in BCIs. On this account, we have focused on the detection of steady-state visual evoked potentials (SSVEPs) and proposed a new detection method that slightly outperforms state-of-the-art competing methods. Given the stereo-vision requirements of immersive applications, we have also explored and compared different stimuli presentation methods. Additionally, we have tackled the problem of detecting interaction error-related potentials due to their possible integration within BCI systems. Since reliable classification of these potentials typically requires long training sessions, special focus has been laid upon the classifier transferability problem. Application-specific improvements were also discussed for immersive robotic embodiment systems, where contextual and adaptive BCIs were proposed as a way to overcome the bandwidth limitations of current BCIs. Recognizing the importance of user intention recognition for effective interface self-adaptations, we have focused on developing a user-agnostic Bayesian framework to track and infer hidden user target goals within the context of navigation. A variety of experiments with human subjects were conducted to evaluate and numerically validate the proposed methodologies and algorithms.