Chiba University engineers designed a disturbance observer around the non-minimum-phase quirk of a 103-gram flapping-wing robot, cutting its worst-axis position error by 53 per cent.

A flapping-wing micro aerial vehicle has obvious appeal and one dominant weakness. It is light, quiet and capable of manoeuvring in spaces a quadrotor cannot, and being light is exactly what makes a gust of wind able to throw it across a room.

A study in Control Engineering Practice from Chiba University takes a specific run at that problem, and the interesting part is not the controller but the thing the controller had to accommodate: these robots move the wrong way first.

The work was done by Abner Asignacion, a specially appointed assistant professor in the Graduate School of Engineering at Chiba University, with Satoshi Suzuki.

The backwards step

Tested on a commercial flapping-wing robot, the team characterised a behaviour known in control theory as non-minimum-phase dynamics. Commanded to move in one horizontal direction, the robot first moved slightly the opposite way before correcting and travelling where it was told.

The physical reason is reasonably intuitive. A flapping-wing craft steers by tilting its whole body, and tilting requires generating a torque — which, for a brief moment, pushes the body in the direction opposite to the eventual translation. The vehicle has to lean before it can go, and leaning starts by shifting the wrong way.

The control consequence is less intuitive and more consequential. A feedback controller watching position sees the robot moving away from where it was asked to go and responds by pushing harder — which makes the initial backwards excursion worse before the correction arrives. The faster the controller reacts, the more it amplifies the very motion it is trying to suppress, and at some gain the system oscillates and goes unstable.

This is a hard limit rather than a tuning difficulty. It means there is a maximum speed at which disturbance correction can be applied, set by the physics of the airframe, and no amount of better software raises it.

Designing around it

The response was a bandwidth-constrained disturbance observer — an estimator that infers the external forces acting on the vehicle from how its motion departs from what the commands predict, built with an explicit ceiling on how quickly it is allowed to respond.

Characterisation came first. The robot was given back-and-forth position commands at frequencies from 0.08 hertz, roughly one cycle every twelve seconds, up to 0.8 hertz, and the non-minimum-phase behaviour proved strongest along the X-axis.

The tuning result is the one that validates the approach. An intermediate observer response gave the best balance: slower settings stayed stable but corrected too little, and faster settings corrected more but oscillated. An optimum in the middle, with failure modes on both sides, is what a genuine physical trade-off looks like — as opposed to a parameter where more is simply better and the engineering is to find the maximum.

On the worst axis, position error fell by 53.1 per cent. Overall three-dimensional position error fell by about 28 per cent.

The gap between those two figures is informative rather than disappointing. The method targets a problem that is specific to the axis where the non-minimum-phase dynamics are strongest, so the largest improvement appears there and the average across all three axes is diluted by the axes that were less affected to begin with.

What the authors expect it to be for

“The control method proposed in this study enables FW-MAVs to fly more stably, even in environments subject to disturbances,” Asignacion said.

“Future applications are anticipated in areas difficult for humans to access, such as infrastructure inspection in confined spaces, search-and-rescue operations at disaster sites, environmental monitoring, and equipment inspections within factories and buildings,” he said.

Those applications are the standard case for flapping-wing flight, and the reasoning behind it is sound. A rotorcraft small enough to enter a collapsed building generates a downwash that stirs up dust and is loud enough to drown out a voice; a flapping wing is quieter and can be built to tolerate brushing against a surface. Confined, cluttered spaces are also where air moves unpredictably, which is why disturbance rejection is the enabling capability rather than a refinement.

What a laboratory test does and does not show

The release leaves one important thing unstated: what disturbance was actually applied during the validation flights.

The 0.08 to 0.8 hertz commands described are position commands used to characterise the robot’s response, not wind. Whether the 53.1 per cent and 28 per cent improvements were measured against a fan, a gust generator, or simply against the vehicle’s own tracking error during commanded manoeuvres is not reported — and the three figures mean quite different things depending on the answer. A controller that tracks a moving reference better is a useful result; a controller that holds position against real wind is the claim the framing invites.

Scale is also a question the work does not address. The test vehicle weighs 103 grams, and the dynamics being compensated depend on the ratio between the torque available and the vehicle’s inertia, both of which change with size. A method tuned to one commercial airframe is a demonstration on that airframe.

Nor is endurance mentioned, and it is the constraint that governs whether any of the anticipated applications is reachable. Flapping-wing vehicles of this mass typically fly for a handful of minutes, which is the real obstacle to inspecting infrastructure or searching a disaster site — better gust rejection makes the minutes more useful without making them more numerous.

The paper appeared online on 4 June 2026 and is scheduled for publication on 1 October 2026 in volume 175. The research was supported by the New Energy and Industrial Technology Development Organization under project JPNP22002 and by JSPS KAKENHI grant JP23K13274. The authors state they have no known competing financial interests or personal relationships that could have influenced the work.

---

Notes

1. “Bandwidth-constrained disturbance observer design for position control of flapping-wing MAVs with non-minimum-phase dynamics.” Control Engineering Practice, volume 175, 2026. DOI: 10.1016/j.conengprac.2026.107099.
2. Chiba University. “Keeping flapping-wing robots stable during disturbances.” EurekAlert!, 4 August 2026. https://www.eurekalert.org/news-releases/1138340.

Photo by Alex Knight on Pexels.com