Advances in language models open up opportunities for the creation of intelligent robot-control systems capable of interpreting general, often vaguely formulated, human target commands and shaping robot behavior based on them under given conditions. The paper proposes a system for the reasoned control of the robot’s behavior, which is the result of logical processing of goal-setting instructions using external information from the robot’s environment, taking into account feedback from the operator. Experiments were conducted to evaluate the success rate of a system of single- and multiagent approaches to forming reasoning and controlling robot behavior. The evaluation results were obtained on a set of reasoning and instructional Large Language Models, such as Deepseek, Gemini, GPT, Mistral, and Qwen. The analysis of the results demonstrates that, in general, the proposed agent-based approaches are capable of effectively controlling the robot’s behavior. The developed agents, built on the basis of the most productive Large Language Models, achieve a success rate of up to 74%, with each approach incorporating its own hallucination reduction mechanism.
Skorokhodov et al. (Sat,) studied this question.