A drone that wants to follow another usually watches it with a camera or receives position information through GPS sharing or a radio link. In darkness, fog, smoke, or places where communication is unavailable, that relative-positioning method can fail.
A team at Duke University's General Robotics Lab asked a different question: what if the drone just listened? Every quadcopter screams, a constant, structured whine from four spinning rotors. Their system, called SonicFly, treats that racket as a homing signal, and chases a target drone through the open air using sound alone.
It sounds almost too simple, and for years the idea was dismissed: a drone is far too loud, in its own ears, to pick out anything else over its rotors. The surprise in this work is that the noise is not only a nuisance to be filtered away. Handled right, it is the signal.
Here is what happened
Pursuit by ear, outdoors. In an August 2026 preprint, the Duke group showed a drone carrying a four-microphone array that used the flight sound of a second drone to estimate its bearing and range, then follow it. Across outdoor tests, it recorded a mean distance-maintenance error of 1.34 metres without vision, inter-robot communication, GPS sharing, or external sensing infrastructure.
The hard trick is ignoring itself. The chasing drone is deafeningly loud in its own ears; its four rotors drown out the faint sound of the target. The system leans on "rotorcraft-informed" acoustic models and a neural estimator that learns to pull the target's signature out of the drone's own din.
A named new idea. The researchers call it "embodied passive aeroacoustic perception": passive because it emits nothing and only listens, embodied because the sensing is built into a flying robot that acts on what it hears.
Light enough to run on the drone itself. The whole kit is deliberately small: four cheap microphones and a compact model that runs onboard, rather than a heavy camera, a radar, or a link back to a base station. That is what makes it plausible on an ordinary drone, not just in a paper.
How it works
The target gives itself away. A quadcopter's rotors make a specific, repeating sound that shifts with distance and angle. Four microphones hear that sound slightly differently, and the tiny timing and volume gaps between them reveal a direction and a rough range, the same way two ears let you point at a sound with your eyes closed. Move the target, and that pattern shifts in a predictable way, which is what lets the drone not just point at it but judge how far off it is.
A neural net does the untangling. Raw drone audio is a mess of wind, echoes, and the listener's own rotors. A trained network, plus a "confidence gate" that ignores uncertain readings, turns that mess into a usable direction-and-distance estimate many times a second.
Passive means lightweight and radio-silent. Because the sensor only listens, it does not reveal itself through an active radar or coordination transmitter and does not depend on an RF link that can be conventionally jammed. It can still be disrupted by wind, acoustic masking, echoes, or deliberate sound spoofing.
Why it matters
It can complement vision and shared positioning. Darkness, haze, fog, and communication loss can degrade cameras or coordinated GPS sharing while rotor sound remains available. The paper demonstrated outdoor pursuit across varied conditions, but it did not test collapsed buildings or every form of smoke, dust, and clutter.
It suggests a counter-drone application. A mobile robot that passively detects and follows another drone could interest security teams. The study demonstrated cooperative leader-follower pursuit, not interception of a hostile or evasive aircraft, so counter-drone use remains a future application rather than a reported result.
Sound is an under-used sense for robots. Almost all robot perception is visual. Showing that a machine can act, not just detect, on sound alone points to cheaper, tougher sensing across robotics, from drones to ground robots working in the dark.
It could matter in contested environments. A lightweight sensing channel that does not depend on a camera or shared radio data could support formation flight or target tracking when communications are degraded. Military relevance is plausible, but the present evidence is a research demonstration, not an operational defence system.
The honest catch
This is a clever proof of concept from one lab, and it is early.
A single, fairly clean setting. The results are outdoor trials tracking one target. Cities are far louder and messier, with traffic, crowds, and many machines at once, and it is unproven there.
Range, clutter, and acoustic similarity remain open. The experiments used a two-blade leader and three-blade follower, creating partially separable sound signatures. Performance may fall when several similar drones, traffic, machinery, echoes, or deliberate masking sounds occupy the same acoustic scene.
Preprint, not a product. The work is a single-lab paper that has not yet cleared full peer review, and no product uses it.
EDITOR'S TAKE
Robotics has leaned heavily on cameras, but the physical world also produces sound that can remain useful when visibility or communication degrades. SonicFly demonstrates that a flying robot can recover relative position from another drone's rotor acoustics while both are in motion. The result is genuinely new, but it used deliberately distinguishable propeller configurations and cooperative outdoor trials. Watch whether the method survives several similar drones, city noise, echoes, and deliberate acoustic interference. If it does, hearing could become a valuable backup channel rather than a replacement for cameras, LiDAR, radar, or GPS.
Quick questions
How can a drone hear anything over its own rotors?
That is the core problem, and the clever part. The chasing drone's own four rotors are far louder in its microphones than the distant target. The team built acoustic models specific to how rotorcraft sound, then trained a neural network to separate the target's propeller signature from the drone's own noise, and added a filter that throws out low-confidence guesses. It is the audio version of picking out one voice in a crowded, noisy room, done fast enough to fly on.
Why not just use a camera or radar?
Each has a different operating envelope. Cameras degrade in darkness, fog, and haze. Radar and radio-based coordination require active signals or links. Passive audio needs only a small microphone array and can provide a complementary channel when vision or communication is unreliable. It is not immune to interference: wind, echoes, acoustic clutter, and spoofing can reduce performance.
Is this meant for teamwork or for stopping other drones?
Both are possible future directions. Following another drone by sound could support radio-silent formation or provide a backup sensor in poor visibility; a defensive robot might also use it to pursue another aircraft. The researchers demonstrated cooperative pursuit, so search-and-rescue and counter-drone missions still require separate testing.
Sources
Embodied Passive Aeroacoustic Perception Enables Relative Sensing and Pursuit Between Aerial Robots (arXiv): the SonicFly paper, with the method and the 1.34 m pursuit result.
IEEE Spectrum: independent coverage of acoustic drone sensing and the Duke work.
Duke General Robotics Lab: the group behind the system and its wider robot-perception research.
Frontier Signal explains frontier technology in plain English. Company and study figures should be independently verified. This is general information, not professional advice.
