Path Planning an‎d Safe Control of Multi-Agent Robotic System Using Graph Neural Networks

2/20/2026 12:00:00 AM

Multi-agent systems enable coordination an‎d decentralized cooperation between autonomous agents, which in turn makes performing complex tasks possible. At the same time, these systems face challenges such as safety, scalability an‎d generalizability to differing environments. This research presents a principled framework for designing secure an‎d scalable controllers for multi-agent systems that employs imitation-learning methods using graph neural networks, reinforcement learning, an‎d control barrier functions. In the presented framework, by taking advantage of the ability of graph neural networks to learn complex relationships in graphs, distributed controllers an‎d control barrier functions are trained for use in multi-agent systems, an‎d by combining this approach with reinforcement learning to design optimal controllers without the need for a model, secure controllers with generalizability to a different number of agents an‎d environments with multiple obstacles are obtained. Simulation results on a multi-agent system consisting of drones confirm the success of this method in achieving high safety an‎d success rates, indicating the promise of this path for achieving distributed an‎d secure solutions for controlling multi-agent systems in complex environments in a model-free manner.

سامانه رباتي چندعامله , شبكه عصبي گراف , مسيريابي , يادگيري تقويتي

Multi-robot system , Graph neural network , Path planning , Reinforcement learning

Ali Cheraghi

Dr. Saeed Shamaghdari