A quadcopter at Duke University can find and follow another quadcopter using nothing but the sound its rotors make. No GPS sharing between the two aircraft, no radio link, no camera, no ground infrastructure.

The system is called SonicFly, and a team at Duke’s General Robotics Lab posted it to arXiv on August 1, 2026, with a project page and flight footage following two days later. Across outdoor pursuit runs in darkness, fog, direct sun and overcast, the follower held a target separation of 11.5 feet (3.5 meters) with a mean distance error of 4.4 feet (1.34 meters). DroneXL has been covering acoustic drone detection since early 2024, and this is the first working system we have seen that puts the microphones on an aircraft that is itself flying and making noise.

The array weighs 10 grams.

SonicFly Separates The Leader’s Rotor Noise From The Follower’s Own

The hard part is not hearing a drone. It is hearing one while flying another. Duke’s follower sits inside its own acoustic field, so the team gave the two aircraft different propellers: two blades on the leader, three on the follower. That harmonic gap is what makes the leader audible.

Multirotor sound is not featureless hiss. Blade passage produces a fundamental tone set by blade count times motor RPM divided by 60, with harmonics stacked above it, so a listener knows roughly where in the spectrum to look. The follower carries four TDK InvenSense MEMS microphones on an 83.5 mm square board with 75 mm spacing, sampling at 12 kHz. A Jetson Orin Nano turns those channels into spectrograms plus inter-microphone phase and level differences, runs a convolutional network that outputs the leader’s bearing and range, gates out low-confidence estimates, and filters the rest.

The useful band turned out to be narrow. The network did best between 1 and 2 kHz, and feeding it the full 0 to 6 kHz range made results worse. Low frequencies carry the most rotor energy, but their long wavelengths give a 75 mm array almost nothing to work with directionally. Higher frequencies resolve direction more finely, then fade with distance and start wrapping phase.

Onboard alternatives to satellite positioning are a running thread here, from vision systems that map their own way through GPS-denied airspace to swarms built to hold formation when the radio link dies. Sound is the channel a radio jammer cannot reach.

The Acoustic Payload Weighs 10 Grams And Draws One Watt

Size, weight and power is where SonicFly makes its clearest case. The microphone board occupies 128 cubic centimeters, weighs 10 grams and draws 1 watt. Duke measured an Intel RealSense D455 at 75 grams and 3.46 watts, and a Livox Mid-360 LiDAR at 265 grams and 6.5 watts.

Both aircraft are Holybro X500 V2 quadrotors on Pixhawk 6X controllers with RTK positioning. The RTK data never reached the perception stack: it labeled the training set and scored the results afterward, and the estimator saw none of it. The follower got BLHeli_32 speed controllers and 900 kV motors to haul the extra load. The 10-gram figure covers the microphone board, not the Jetson bolted on beside it.

The Tracking Numbers Show A Proof Of Concept, Not A Product

Bearing is where the system struggles. After confidence gating and filtering, mean absolute bearing error was 31.2 degrees, improved from 45.7 degrees raw. Range held up better at 1.49 meters RMSE. The follower stayed inside the desired lateral corridor for 78% of total flight time across trials.

Training and testing ran at separations of 4.9 to 23 feet (1.5 to 7 meters), the follower flying about 20 inches above the leader to stay out of its wake. Error did not fall off smoothly with distance. It was lowest at intermediate separations, because up close the leader’s near-field wake scrambles the signal, and far out the harmonics sink into the follower’s own broadband roar.

The contrast with stationary listening is stark. Twelve ground units spread across a 98 by 66 foot (30 by 20 meter) field resolved bearing to 6.86 degrees and range to 1.26 meters, roughly five times better than the flying array managed. That gap is the whole difficulty of the problem in one number. The Duke authors are direct about it, describing a system that demonstrates pursuit and relative-state maintenance rather than precise formation control. This is also an arXiv preprint rather than peer-reviewed work, and the figures come from the team that built the thing.

Acoustics Outperformed Cameras And LiDAR In Degraded Conditions

Duke scored all three modalities on the same flight data, gathered in evening light and haze, counting the share of time each could locate the leader within 5 meters. Acoustics managed 76.57%. LiDAR reached 61.13%. Vision came last at 52.48%. Fog and darkness cost a microphone nothing.

The flight campaign spanned 46 to 90 degrees Fahrenheit (8 to 32 Celsius), humidity from 19% to 93%, and winds up to 23 mph (37 km/h), all of which bend and scatter sound in ways a lab bench never shows. Leader paths included straight lines, C-curves, L-turns and, because graduate students are graduate students, a route that spelled out “DUKE.”

Ground-based acoustic detection is already a commercial category rather than a laboratory curiosity, from the cheap Ukrainian microphone arrays that spawned North American imitators to the sensor networks strung across Ukrainian towns to call Shahed inbounds and Belarusian systems on the same principle. What none of those do is fly.

DroneXL’s Take

This story is going to get read as counter-drone technology, and that read is premature. A follower that tracks a leader by ear sounds like an interceptor, right up until you notice that the Duke team engineered the target to be audible in the first place.

Two-blade propellers on the leader, three on the follower, picked so their harmonics would separate. A hostile drone will not accommodate you like that, and the paper says plainly that this work covers a single leader-follower pair. Whether the same network can hear an aircraft it was never trained on, one whose blade count nobody selected for contrast, is not answered anywhere in the results.

I read the real result differently. The interesting number is 10 grams and 1 watt for a sensor that works in fog and darkness without line of sight, at a fraction of what a LiDAR puck costs in payload. That is not a weapon. That is redundancy for aircraft flying in formation, and formation is what falls apart first when GPS goes away and the radio link gets stepped on. Ukraine has spent three years proving how routine both conditions have become.

The counterpoint deserves a hearing, because ground-based acoustic detection has earned its place: cheap microphone arrays are doing real counter-drone work today, and a listener that ignores weather and daylight is obviously attractive to that market. Fine. But 31.2 degrees of bearing error inside a 23-foot envelope is not a detection system, and dressing it up as one does the researchers a disservice.

The version that matters is the one where the blade-count crutch comes out and a follower tracks an unmodified consumer quadcopter it has never heard before. Publish that, and the counter-drone read becomes fair. Until then what stands is that a flying robot can navigate by the noise its neighbor makes, which is enough. Universities keep producing the drone research that moves this field, as Houston’s work on formally provable flight behavior showed last month, and the industry keeps skipping the limitations section to write the press release.

Sources: Duke University General Robotics Lab, arXiv:2608.00401, IEEE Spectrum

DroneXL uses automated tools to support research and source retrieval. All reporting and editorial perspectives are by Haye Kesteloo.