Robust Multi-Agent Reinforcement Learning for Small UAS Separation Assurance under GPS Degradation and Spoofing

A team from the University of Texas at Austin, led by Alex Zongo and Peng Wei, has developed a breakthrough Multi-Agent Reinforcement Learning (MARL) framework that ensures safe separation for drone swarms operating under GPS degradation and spoofing attacks. The research addresses a critical vulnerability in cooperative surveillance systems where each drone broadcasts its GPS-derived position—when these signals are corrupted, the entire air traffic state becomes unreliable. The team's key innovation is casting this problem as a zero-sum game between the drones and an adversarial attacker.

The researchers derived a closed-form mathematical expression for the worst-case adversarial perturbation that corrupts observed states, bypassing the need for computationally expensive adversarial training entirely. This enables linear-time evaluation in state dimension while approximating the true worst-case perturbation with second-order accuracy. They proved that safety performance degrades at most linearly with corruption probability under Kullback-Leibler regularization, providing theoretical guarantees for real-world deployment.

In high-density sUAS simulations, the integrated MARL policy gradient algorithm achieved near-zero collision rates even when 35% of GPS observations were corrupted—significantly outperforming baseline policies trained without adversarial perturbations. The system's robustness comes from its ability to learn counter-policies that anticipate and mitigate the most damaging spoofing attacks, ensuring safe operation in contested environments where GPS signals cannot be trusted.