This paper presents a brain-computer interface (BCI) architecture for robotic devices. Two datasets are used to perform a simulation of real-time classification, which is a pseudo-online technique, to measure the performance of the proposed BCI architecture. Dataset V comprises three mental tasks, namely, left or right hand movement imagery and the thought of generating words beginning with a given random letter and dataset IIIa comprises four motor mental tasks, namely right hand, left hand, foot, and tongue movement imagery. The partial auto-correlation function is used to estimate the size of the time windows. Power spectral density is used for feature extraction and the Kullback-Leibler symmetric divergence is adopted to select the electroencephalogram channel and frequency. A Bayesian classifier is used to recognize the mental tasks and a reclassification model is proposed to improve the response of the classifier. The proposed BCI can identify 3 and 4 mental tasks with accuracies of up to 94.9% and 72%, respectively.