This paper develops an event-related potential (ERP)-based adaptive model for the control of humanoid robot movements in an environment cluttered with obstacles based on live video feedback. This model adaptively determines the repetition number according to an individual’s mental state to speed up the robot control cycle. N200 and P300 potential features increase in the frontal and occipital areas when using robot images as visual stimuli, so it is able to effectively recognize target visual stimuli by processing Fisher’s linear discriminant analysis (FLDA) and to identify a subject’s intention by using support vector machine (SVM), in parallel. The offline evaluations show that, compared with a nonadaptive model, the adaptive model increases the accuracy rate from 88.8% to 92.9%, a change of 4.1%, and the information transfer rate (ITR) from 41.3 to 46.3 bits/min, a change of 5.0 bits/min. Eight subjects participated in telepresence controlling a NAO humanoid robot to move in an office environment cluttered with obstacles. The successful maneuvers demonstrate that the brain-controlled humanoid robot can be applied for surveillance and exploration in unknown environments based on live video feedback, which are evaluated by using new metrics for the performance of the brain–robot interaction (BRI) system.