The researchers asked whether perturbing a gyroscope could be associated with displacement, rather than perturbing a sensor that directly reports position or translational motion. During the simulated attack, the vehicle showed y-axis drift of about 0.05 metres per second while its altitude deviation remained below 5 centimetres.
In an illustrative indoor run, the first four positive-sign triggers began at about 53.8 seconds. They coincided with a peak roll bias of about minus 5.5 degrees and displacement of roughly 1.1 metres toward the negative y direction. The result was approximate and came from that illustrative run.
It is an arXiv preprint labeled arXiv:2608.25319v1 and dated 26 August 2026.
The study varied estimator settings
The study analyzes this approach in a PX4-style flight stack and reports validation through 81 simulation runs and more than 10 indoor and outdoor physical experiments.
The sensitivity analysis used an 81-run extended Kalman filter, or EKF, weighting sweep over a 9 by 9 grid of settings for accelerometer trust and gyro-bias adaptation. In ordinary terms, the estimator turns sensor readings into a working picture of tilt and motion, and the controller uses that picture to decide how to keep the aircraft stable.
The sweep associated higher residual roll bias with larger Rscale and lower residual roll bias with larger Qscale.
A local control model estimated about 0.6 metres of lateral offset for each degree of residual tilt under representative gains of about 0.3 per second in the velocity loop and 1.0 per second in the position loop. The estimate is a local model approximation, not a universal physical scaling law.
Controlled indoor and outdoor flights
Physical flight validation used a Holybro X500 V2 with a Pixhawk 6X and an MPU6500. Indoor position measurements came from VICON at 10 hertz, while outdoor measurements used GNSS at 5 hertz.
The indoor run had 16 phase-aligned triggers out of 34 scheduled windows, or 47 percent. Across 10 repeated indoor trials, the standard deviation of that rate was 5.8 percentage points. The study reports this as descriptive variation rather than a confidence interval.
An outdoor run lasted about 60 seconds, with the reported attack window extending from 7 to 55 seconds and including 16 AGR triggers. The run reported up to roughly 1.5 metres of lateral displacement along with motion on the x-axis. The motion estimate used fused-state vehicle odometry without independent ground truth.
A broader sensor check
A separate benchmark evaluated 10 drone-grade inertial sensor models, including six not covered by the cited prior acoustic-injection datasets. Several sensors showed susceptibility under speaker-detached excitation, with the strongest coupling usually reported on roll and pitch rather than yaw.
For the BS BMI088 entry, Dmax was listed as 1.52 metres, with the x and y axes shown as affected. Dmax is a thresholded sensor-disturbance standoff measure from a constrained test. It is not a measurement of drone displacement or a fundamental acoustic range.
What the study leaves open
The evidence is limited to a PX4-style control model, simulations, controlled indoor and outdoor UAV experiments, and constrained PX4 EKF-in-the-loop sensor tests. It does not establish that all UAVs, inertial sensors or controller settings are susceptible.
Fully detached, black-box, moving-target free flight was not demonstrated. That leaves open how well the pathway would transfer to other autopilot stacks, sensor installations, changing geometry or faster motion.
Paper data and sources
Original title: SonicNudge: Controlled Displacement of Hovering UAVs via Estimator-Controller Coupling
Authors: Shaocheng Luo, Ashir Raza, Haocheng Meng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text