Introduction

Unmanned Aerial Vehicles (UAVs) play pivotal roles today in photography, deliveries, rescue missions, border surveillance, and military operations. However, the growing availability of signal disruption tools (GPS spoofing, gyroscope jamming, magnetometer manipulation) poses significant threats to autonomous systems. Even a slight navigational drift can turn a mission into a disaster.

Why Physical-Attack Robustness Matters

Traditional safe RL methods or adversarial trainings rely on known attack scenarios. In practice, it’s impossible to anticipate every possible manipulation—an adversary could employ novel jamming or optical disruption techniques. Iterative adversarial training is computationally expensive and often poorly generalizes to unseen scenarios.

How ARMOR Works

ARMOR (Adaptive Robust Manipulation-Optimized State Representations) is a model-free RL controller that learns a robust latent representation of the UAV state in two stages:

  1. Teacher encoder: a Variational Autoencoder (VAE) trained with privileged attack information (GPS bias, jamming duration) producing an attack-aware latent vector z_t.
  2. Student encoder: a temporal VAE with LSTM, relying solely on raw sensor-read history, mimics z_t, classifies attack type, and reconstructs inputs.

After training, the RL policy uses only the student-generated vectors, enabling deployment without privileged data.

Real-Life Scenarios

  • Ghost Dragon (Ukraine): autonomous quadcopters using optical navigation bypass GPS jamming by tracking unique terrain features.
  • GPS Jamming in the Baltic Region: disruptions around Kaliningrad affected commercial flights in Poland and Estonia, forcing reroutes and delays.

Conclusion

ARMOR combines latent representations with knowledge transfer between teacher and student encoders, achieving zero-shot robustness to unseen attacks with lower training cost. It’s a milestone for autonomous UAV systems.

📎 Links