摘要
Previous studies have demonstrated the possibility of using functional MRI to control a robot arm through a brain-machine interface by directly coupling haemodynamic activity in the sensory-motor cortex to the position of two axes. Here, we extend this work by implementing interaction at a more abstract level, whereby imagined actions deliver structured commands to a robot arm guided by a machine vision system. Rather than extracting signals from a small number of pre-selected regions, the proposed system adaptively determines at individual level how to map representative brain areas to the input nodes of a classifier network. In this initial study, a median action recognition accuracy of 90%was attained on five volunteers performing a game consisting of collecting randomly positioned coloured pawns and placing them into cups. The 鈥減awn鈥?and 鈥渃up鈥?instructions were imparted through four mental imaginery tasks, linked to robot arm actions by a state machine. With the current implementation in MatLab language the median action recognition time was 24.3 s and the robot execution time was 17.7 s. We demonstrate the notion of combining haemodynamic brain-machine interfacing with computer vision to implement interaction at the level of high-level commands rather than individual movements, which may find application in future fMRI approaches relevant to brain-lesioned patients, and provide source code supporting further work on larger command sets and real-time processing.