The problem of interpreting operator control commands, expressed in natural language, in a formalized form when controlling a robotic agent in a complex world seems very relevant. This work compares several methods based on neural network models of the Transformer architecture to obtain a formalized representation of user commands on choosing the direction of movement. Neural network models, depending on the chosen method, are configured to solve one of the following tasks: multi-label multi-class classification, named entity recognition (command attributes) and generation of text sequences (structured commands). The feasibility of using Few-Shot prompting on a large trained generative model is also tested. The models were fine-tuned using a training corpus of Russian text commands and their formalized representations. The corpus was formed in a semi-automatic mode by generating commands according to given templates using dictionaries of synonyms for actions characterizing movement commands and attributes characterizing objects in a given world. Testing of all proposed models was carried out on a manually assembled and labeled corpus. Based on the evaluation results on the test corpus, the best model was selected on the basis of a set of metrics, and it was a model based on the task of generating text sequences. The final model was implemented into a robot control system in a virtual environment based on the Gazebo simulator. Code available at https://github.com/sag111/commands-dataset-ru .