MANAGEMENT OF MULTI-AGENT ROBOTIC SYSTEMS USING ARTIFICIAL INTELLIGENCE TECHNOLOGIES

This article addresses the problem of managing multi-agent robotic systems using artificial intelligence technologies. A hybrid model combining the Vision-Language-Action (VLA) architecture with the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm is proposed for the synergetic control of a group of robots. The model is built on the Centralized Training with Decentralized Execution (CTDE) principle, with a Social-Force component incorporated into the reward function to automatically maintain a safe distance between agents. Experiments were conducted in the NVIDIA Isaac Sim environment using a dataset of over 10,000 trajectories and scenarios involving 4 to 16 agents. The results show a 90% task success rate, a 0.5% collision probability, and 89% energy efficiency, along with a significant reduction in communication load compared to a rule-based method and standard MADDPG. These findings confirm that integrating VLA models with multi-agent reinforcement learning improves the adaptability, fault tolerance, and collective decision-making capability of robotic swarms, opening prospects for application in agriculture, logistics, and smart-city infrastructure.