Practical Robust Formation Control for Nonlinear Multiagent Systems via Generative Adversarial Learning Framework: Theory and Experiment

Abstract

Cyber attacks and disturbances greatly impair the performance of formation tasks in multiagent systems (MASs). To achieve robust formation control against these challenges, this article proposes a generative adversarial learning framework that is theoretically transparent and practically applicable. Rather than relying on an end-to-end deep neural networks (DNNs) architecture, our work leverage a double robust structure that combine the representation capabilities of DNNs with established, theoretically grounded linear control theory, ultimately achieving a practical, learning-based robust formation for MASs. Initially, generative adversarial networks (GANs) are used to linearize agent dynamics under false data injection (FDI) attacks and external disturbances. Subsequently, a proportional-integral (PI) protocol is employed to achieve overall robust formation. We present rigorous theoretical analyses of both stages, demonstrating the guaranteed convergence of GANs training and the closed-loop formation errors. Our approach is directly validated through a series of physical experiments involving multi-quadrotors, demonstrating robustness against attacks and disturbances during formation flights, without the sim-to-real gap commonly encountered in learning-based control frameworks.