Control and Coordination of Multi-Drone Systems Using Graph Neural Networks

Multi-robot systems show great promise in performing complex tasks in areas ranging from search and rescue to interplanetary exploration. Yet controlling and coordinating the behaviors of these robots effectively is an open research problem. This research investigates techniques to control a multi-drone system where the drones learn to act in a physics-based simulator using demonstrations from artificially generated motion data that simulate flocking behavior in biological swarms. Using these demonstrations enables faster training than approaches where the agents start learning from scratch. The Graph Neural Network (GNN) controller used for the drones learns an efficient representation of low-level interactions in the system, allowing the proposed method to scale to more agents than in training data. This work also discusses techniques to improve performance in the face of real-world challenges such as sensor noise.