Spectral Characterization of UAV Attitude Degradation Under Real-Time Vision Workloads: A SITL Framework for Pre-Deployment Diagnosis

Quispe-Checca E.; Sacoto-Cabrera E.J.; Laurente G.E.G.; Mego L.W.U.; Silva-Alvarado V.L.; Lloret J.; Moreno-Cardenas E. · IEEE Access (2026) · DOI: 10.1109/access.2026.3722205

Resumen

Real-time vision inference executed concurrently with unmanned aerial vehicle (UAV) flight control can introduce computational delays that couple with the attitude control loop and degrade dynamic stability. This work presents a spectral characterization of the disturbance induced by embedded vision workloads on a quadrotor controlled by cascaded Proportional-Integral-Derivative (PID) loops. A Software-In-The-Loop (SITL) platform based on PX4 Autopilot, Gazebo Classic 11, and ROS2 Humble was employed. A YOLO26n detector was used exclusively as a controlled computational stressor, generating latency regimes of 2.5 ms with GPU acceleration and 84.8–90.7 ms under CPU inference. Experimental results show that CPU inference induces a narrowband elevation in the power spectral density (PSD) of roll and pitch concentrated in the 1–3 Hz band. At 2.0 m altitude, pitch PSD increased by 9.07× (+10.6 dB), while under in-ground-effect (IGE) conditions at 1.5 m the disturbance decreased to 1.82× (+5.9 dB). A continuous altitude sweep revealed a maximum amplification of 698× near the IGE/out-of-ground-effect transition when the camera pipeline remained coupled to the simulator physics loop. Architectural isolation experiments show that the dominant degradation mechanism is not the latency magnitude but the temporal coupling of the inference pipeline to the dynamics integrator: a decoupled pipeline with higher median latency produced no spectral elevation. An analytical critical latency threshold of λcrit ≈ 76 ms was derived from the delay margin of the outer position loop; this threshold only becomes active when the pipeline is coupled to the control loop. A closed-loop sensitivity analysis predicts a resonance peak inside the 1–3 Hz band, consistent with the measurements. Finally, three quantitative pre-deployment indicators are proposed as SITL-derived reference values for validating real-time vision systems integrated into PID-controlled UAV platforms operating in controlled low-wind environments such as greenhouses.

Instituciones: Universidad Nacional de San Antonio Abad del Cusco