Blind Spot in Sensor Fusion: Exploiting an Architectural Vulnerability to Hijack and Precisely Control UAVs

To achieve nimble real-time control while remaining robust to error accumulation, PX4 and ArduPilot UAV control pipelines typically combine high-rate IMU-based attitude controls with low-rate GPS-aided state estimation. We show that this foundational design choice introduces a persistent blind spot: injected gyroscope bias immediately alters control outputs yet remains invisible to the state estimator until slow GPS position updates reveal state inconsistencies. Leveraging this architectural vulnerability, we design a black-box reinforcement learning framework that learns to induce trajectory manipulation. Through asymmetric actor-critic training, we derive a highly effective attack policy that, given bias-injection and external observation capabilities, achieves over 85 % success in steering UAVs into a 10 m-radius cylindrical target region, transfers across different autopilot and airframe types without retraining, bypasses a state-of-the-art anomaly detector and built-in failsafes, and transfers to real-world UAVs.