Learning Resilient Formation Control of Drones with Graph Attention Network

TL;DR

This work tackles resilient formation control for multidrone swarms under communication loss and adversarial DoS attacks. It introduces a graph attention network–based controller that dynamically weights inter-agent information and learns a dual-mode policy combining leader-follower and distributed strategies. Through extensive Unity-based simulations and zero-shot Sim2Real experiments with five drones, the approach outperforms traditional formation controllers and DRL baselines in formation accuracy and collision avoidance, even under DoS conditions. The results demonstrate practical viability for robust, scalable drone formations in cluttered or contested environments.

Abstract

The rapid advancement of drone technology has significantly impacted various sectors, including search and rescue, environmental surveillance, and industrial inspection. Multidrone systems offer notable advantages such as enhanced efficiency, scalability, and redundancy over single-drone operations. Despite these benefits, ensuring resilient formation control in dynamic and adversarial environments, such as under communication loss or cyberattacks, remains a significant challenge. Classical approaches to resilient formation control, while effective in certain scenarios, often struggle with complex modeling and the curse of dimensionality, particularly as the number of agents increases. This paper proposes a novel, learning-based formation control for enhancing the adaptability and resilience of multidrone formations using graph attention networks (GATs). By leveraging GAT's dynamic capabilities to extract internode relationships based on the attention mechanism, this GAT-based formation controller significantly improves the robustness of drone formations against various threats, such as Denial of Service (DoS) attacks. Our approach not only improves formation performance in normal conditions but also ensures the resilience of multidrone systems in variable and adversarial environments. Extensive simulation results demonstrate the superior performance of our method over baseline formation controllers. Furthermore, the physical experiments validate the effectiveness of the trained control policy in real-world flights.

Figures (14)

• Figure 1: Illustration of one direct graph definition.
• Figure 2: The drone formation with the DoS attacker.
• Figure 3: The overall architecture of the proposed distributed resilient formation control with GAT. The observable state of the attacked agent $\boldsymbol{s}'_2$ is extracted from a GAT with neighbour agents' states. If the agent 2 is within the attack range, the information $\boldsymbol{p}_{12}$ is not available. The RL is trained with the extracted state representation $\boldsymbol{s}'_2$.
• Figure 4: The overall architecture of the designed policy NN.
• Figure 5: The simulation environment developed in Unity. The information $\boldsymbol{p}_{1a}$ of adversarial follower $F_2$ is blocked if it is within the attack range.