The paper proposes an approach to and solves the problem of controlling a robot by using neural interface technology, describes the general scheme and working principle of the main idea of non-invasive neural interface control of a robot using the original convolutional neural network. The authors describe the principles of an original convolutional neural network and an approach to the modern network design, present a model of a one-dimensional convolutional network based on the principles of a human inner ear. The structure of a software package is proposed. The results of a comparison of algorithms for the analysis of human brain evoked potentials used in the design of brain-computer interfaces are presented. The authors used the Fourier transform algorithm and the multidimensional synchronization index (MSI) algorithm in various modifications to perform the experiment. Analysis of the initial signal, the accumulated evoked potential, in addition to the accumulated evoked potential spectrum were proposed as variations. Linear correlation was also evaluated with analysis using a user-derived reference signal sample and various variations of wavelet filtering. In addition, model signals, which were a combination of white noise and a harmonic oscillation simulating a stable visual evoked potential, were used. The best results (error rate <10%) with an analysis time of 3 s were obtained for the MSI of the original signal, MSI with the Fourier transform. Also in this list, there is a MSI where the wavelet filtering result of coherent accumulation was used as an etalon, a linear correlation coefficient. In a MSI the evoked potential, recovered after the wavelet transform, was used as an etalon.