Learning Resilient Formation Control of Drones With Graph Attention Network

Multidrone systems offer notable advantages in various missions, such as search and rescue, environmental surveillance, and industrial inspection, providing enhanced efficiency and redundancy over single-drone operations. However, ensuring resilient multidrone formation in dynamic and adversarial environments, such as during communication loss or cyberattacks, remains a significant challenge. Traditional approaches often struggle with complex modeling requirements and scalability issues. Among them, leader-follower methods rely heavily on predefined hierarchies, making them vulnerable to single-point failures, while distributed methods incur high communication costs and lack efficient mechanisms to dynamically adapt to changing environments. This article proposes a novel learning-based formation control method to enhance the scalability and resilience of multidrone formations. First, a graph attention network (GAT) is leveraged to dynamically model interagent relationships and prioritize critical interactions among variable neighbors via attention mechanisms with bounded communication overhead. Second, a dual-mode control strategy is designed, integrating leader-follower and distributed control approaches to optimize communication costs while maintaining formation performance. Third, deep reinforcement learning is utilized to train the GAT-based controller, achieving objectives, such as maintaining formation tightness, avoiding collisions, and ensuring resilience against Denial-of-Service (DoS) attacks. Extensive simulations demonstrate superior performance of our method over baseline controllers under normal and adversarial conditions. Furthermore, real-world flight experiments validate the effectiveness and generalizability of the trained policy.