Arquitectura multi-agente descentralizada para detección de eventos en el ambiente

Abstract

Environmental event detection is a key challenge in modern robotics, particularly in systems that require autonomy, cooperation, and distributed decision-making. Decentralized multi-agent systems (MAS) emerge as anefficient alternative to centralized approaches, offering greater scalability, robustness, and fault tolerance. This work proposes the design and implementation of a decentralized multi-agent architecture based on ROS 2 for environmental event detection, using TurtleBot3 Burger mobile platforms and deep reinforcement learning. The research followed an experimental and applied approach, structured into three phases: (1) analysis and selection of physical variables, algorithms, and robotic platform; (2) development of the decentralized architecture integrating a leader agent trained through the Deep Q-Learning (DQN) algorithm and follower agents controlled by the ROBOTIS Follower algorithm; and (3) system validation in the Gazebo simulator using the ROS2middleware for inter-node communication and independent namespace management for each robot. Over 1000 training episodes, the leader agent showed a progressive improvement in cumulative rewards, achieving stable learning and autonomous navigation without collisions. In the multi-agent simulation, the follower robots accurately replicated the leader’s trajectory, maintaining stable formations and effective communication. Finally, physical tests confirmed the successful transfer of the trained model, maintaining consistency with the simulated behaviors. The results confirm that combining deep reinforcement learning with a decentralized ROS 2-based architecture is a viable strategy for developing cooperative behaviors in mobile robotics. The designed system reduces the dependence on a central controller, enhances scalability, and establishes a solid foundation for future validations in more complex physical environments.