This paper presents the results of one of the solutions to the problem of increasing the speed of decision-making by an intelligent agent when modeling the behavior of complex systems on a virtual electronic polygon. Such a training ground is currently considered as an instrumental platform for testing technologies for training intelligent agents in conditions of varying complexity in order to subsequently transfer the developed methods to real objects for solving practical problems. As an example, the control of a robotic device operating in an enclosed space is considered. The article describes the technology of reducing the volume and dimension of the processed data in order to increase the responsiveness to changes in the situation and the development of solutions for moving a robotic device. The technology is based on the preprocessing of video images for the formation of a training sample, as well as the procedure and results of deep learning of a convolutional neural network. The paper uses an open source library of OpenCV computer vision algorithms implemented in C / C++. It is shown that focusing on the selection of object boundaries can significantly reduce the amount of data for analyzing the situation and increase the speed of decision-making by the robot to move.