In November 2026, researchers at ETH Zurich unveiled FlexiQuad, a quadcopter that deforms on impact to absorb the shock. Its defining feature is a ring-shaped frame with no central body, and its takeoff weight is 405 g. According to the research paper, in manually piloted outdoor tests it exceeded 84 km/h vertically and 77 km/h horizontally, and it suffered no damage when it hit a panel head-on at 4.5 m/s (16.2 km/h). In separate research using an improved version of the aircraft, the team also succeeded in experiments where the drone autonomously regained stable flight after hitting a wall. However, the "maximum of about 25 km/h" impact test speed that ETH announced does not match any figure that can be confirmed in the published papers, and both papers are still pre-peer-review. Can a flexible airframe deliver high-speed flight, crash resistance, and post-collision recovery at once? This article examines the potential and the challenges that remain before practical use.
A Ring Frame with Direction-Dependent Stiffness, and Batteries Distributed Across the Motors
FlexiQuad is not the first research on crash-resistant drones. An insect-style drone announced by the Swiss Federal Institute of Technology in Lausanne (EPFL) in 2017 attached a 0.3 mm-thick fiberglass frame to a central case with magnetic joints; on impact, the joints detached to release the shock. Reportedly, it sustained no permanent damage after more than 50 drops from a height of 2 m. Stefano Mintchev took part in that research. FlexiQuad differs from earlier designs in that it eliminates the central case entirely and makes the whole ring-shaped frame connecting the four motors flexible.
FlexiQuad was developed by Luca Girardi of ETH Zurich (now at the ETH Mobile Robotics Lab) and Mintchev, who was a professor at ETH Zurich at the time of development (and is now a professor of agricultural robotics at the University of Bonn). Development took place in ETH Zurich's Environmental Robotics Lab. The university has filed a patent application on the structural design and says the project was nominated for the 2026 Spark Award. According to ETH, Girardi worked on it for about five years. When the first aircraft without a central reinforcement was flown, he doubted whether it could really fly. In the announcement he said, "Soft doesn't automatically mean slow."
ETH's announcement does not name the paper, but a paper that Girardi, Gabriel Maquignaz, and Mintchev submitted to arXiv on November 7, 2025 (arXiv:2511.05426) covers the same aircraft. As of November 4, 2026, only the first version is available, and it is marked as awaiting peer review. According to the paper, the aircraft has a diameter of 216 mm. Its skeleton is a ring made of four strips (each 145.5 × 54 × 0.5 mm) of a fiberglass composite called FR-4, and four 330 mAh 3-cell batteries are placed one under each motor.
This design has two key features. The first is "anisotropy," meaning the frame's stiffness differs depending on the direction of the applied force. The ring is soft in its in-plane direction, while its bending stiffness against the propeller thrust direction is 2.2 times higher. In flight, the stiffness in the thrust-bearing direction keeps the shape; when the drone hits a wall and force is applied in the plane of the ring, it deforms substantially. According to the paper, the overall stiffness of the frame is three orders of magnitude lower than that of conventional aircraft, roughly 1,000 times softer.
The second is that the batteries are distributed under the motors rather than concentrated in the center. If the batteries are gathered in the middle, the soft ring frame has to accelerate while supporting a central weight, which increases vibration. When the research team examined a model with the same frame stiffness but concentrated batteries using finite element analysis, vibration during acceleration increased 13-fold at the motors and 8-fold at the control unit. These are simulation comparisons, not measurements from the real aircraft. The paper also estimates the optimal stiffness range at 0.006–0.77 N/mm, says it can be applied to aircraft of 20–3,000 g, and notes that this overlaps with the stiffness of bird and insect wings.
So can such a soft structure really fly fast? In manually piloted outdoor tests, it recorded linear acceleration above 3 g and angular acceleration above 300 rad/s². The thrust-to-weight ratio was 4.15, and the top speeds exceeded 84 km/h vertically and 77 km/h horizontally. ETH's announcement says "up to 80 km/h" and the paper's abstract says "over 80 km/h," but the figures separating vertical and horizontal speeds appear in the body of the paper. The paper concludes that frame flexibility does not necessarily impair agility, and Girardi explains that the aircraft achieved linear and angular acceleration comparable to stiff drones.
Impact Loads Reduced by Up to 39 Times, but the Numbers Depend on Test Conditions
Different tests report different figures for FlexiQuad's crash resistance. In flight, a head-on collision at 4.5 m/s caused no damage; in bench tests with the aircraft fixed in place, peak load fell by up to a factor of 39; and in simulation, peak deceleration was less than a quarter of that of a rigid frame. The test conditions and comparison targets for each are summarized below.
| Figure | Test type | Conditions | Comparison target |
|---|---|---|---|
| No damage at 4.5 m/s | Flight test | Manually piloted, head-on collision with a vertically mounted MDF panel | None |
| Peak load down by up to 39× | Bench test | A 5 m/s pendulum aluminum rod strikes near the propeller of a fixed aircraft | Rigid carbon aircraft of the same mass |
| Tilting torque down 15× | Bench test | Same as above | Rigid carbon aircraft of the same mass |
| Peak deceleration under 1/4 | 2D finite element analysis | Collisions at 1, 3, and 5 m/s calculated | Rigid frame model |
These results were obtained under different conditions, so the 39-fold reduction in load and the reduction of peak deceleration to under a quarter cannot be compared on the same basis. In particular, the results of actual flight tests, tests with a fixed aircraft, and simulations need to be kept distinct.
In the flight tests, an aircraft accelerated under manual control was flown head-on into a vertically mounted MDF panel at 3 m/s and 4.5 m/s. At both speeds, no structural or functional damage was observed. The peak acceleration experienced by the control unit in the 3 m/s collision was 25.2 g. Because the soft ring frame deforms, the duration of the impact is extended to about 10 times that of a rigid aircraft. By spreading the same impact over a longer time, the peak acceleration transmitted to the control unit is reduced.
The bench tests, on the other hand, do not reproduce collisions in flight. The aircraft was fixed to a stand, and a pendulum aluminum rod struck near the propeller at 5 m/s (kinetic energy 5.06 J), imitating a glancing collision with an obstacle. Compared with a rigid carbon aircraft (220 mm wheelbase) with the same mass and electronics, the peak load was up to 39 times lower and the torque tilting the aircraft was 15 times lower. However, the number of trials was limited: six for FlexiQuad and eight for the rigid aircraft.
The "4 times greater crash resistance" and "no damage in a 5 m/s head-on collision" results that appear in the paper's abstract also did not come from actual flight tests. The research team used the inverse of the peak acceleration on the control unit as an index of crash resistance and evaluated collisions at 1, 3, and 5 m/s with 2D finite element analysis. The result was an estimate that peak deceleration would be less than a quarter of that of a rigid frame. The analysis model itself was validated with drop tests at 0.5–2 m/s, but the highest speed at which a head-on collision in flight is confirmed to have caused no damage is 4.5 m/s.
This brings us to the discrepancy between ETH's official announcement and the paper's figures. ETH says it conducted impact tests at up to about 25 km/h, but the published paper confirms only a 4.5 m/s (16.2 km/h) damage-free head-on collision for the original FlexiQuad and, for the improved aircraft described below, exploratory autonomous recovery tests up to 6.05 m/s (21.8 km/h). 25 km/h is equivalent to about 6.9 m/s, which is above both figures. According to ETH, the impact test speed was determined mainly by the limited space available for acceleration.
That said, this difference alone does not prove that ETH's announcement is wrong. Tests not described in the paper may have been conducted. The conditions also differ: the 4.5 m/s test was a manually piloted outdoor flight, while the 6.05 m/s test was an exploratory trial indoors using motion-capture position assistance. At the least, the two papers reviewed here contain no data supporting a collision at 25 km/h.
REFLEX: Autonomous Recovery in 19 of 22 Wall Impacts
FlexiQuad research is moving beyond surviving collisions toward recovering flight afterward. In a follow-up paper called "REFLEX" (arXiv:2610.15867), submitted to arXiv on October 11, 2026, the authors report that in experiments where the drone hit a wall at an average of 5.03±0.09 m/s, it recovered autonomously in 19 of 22 trials (86%). With the standard PX4 (open-source autopilot software) on the same aircraft under the same conditions, only 1 of 14 trials (7%) succeeded. In REFLEX tests, recovery from the last 15 collisions succeeded consecutively in a single flight. The paper is a pre-peer-review submission to ICRA 2027, a work in which Jakob Hermeler and Luca Girardi contributed equally, supervised by Stefan Leutenegger and Stefano Mintchev.
The experiments used an improved version of FlexiQuad weighing 0.371 kg, different from the original 405 g model. While the flexible ring frame can absorb collision shock, vibration remains after impact, making it hard to stabilize flight. REFLEX therefore uses reinforcement learning to train post-collision recovery maneuvers in simulation. Training used post-impact state data obtained by crashing the real aircraft 47 times. The trained control policy runs on the aircraft's onboard flight controller, but indoor motion capture is used for the timing of starting recovery control, reinitializing velocity estimation, and deciding when to return to normal flight control. Thus, it has not been demonstrated that post-collision recovery can be completed using onboard sensors alone.
Comparison with prior work shows what is distinctive about REFLEX. In the earlier studies summarized in the paper, the speeds at which aircraft autonomously regained flight after hitting a wall ranged from 2.0 m/s in 2017 to 3.7 m/s in 2025. Specifically, Dicker et al. (2017) reached 2.0 m/s, Liu and Karydis (2021) 2.58 m/s, the same authors (2023) 3.5 m/s, and Bredenbeck et al. (2025) 3.7 m/s. By contrast, REFLEX recovered in 19 of 22 trials after impacts at about 5 m/s.
{
"type": "bar",
"title": "Speed of autonomous recovery from wall collisions (per the REFLEX paper)",
"unit": "m/s",
"categories": ["2017 Dicker et al.", "2021 Liu and Karydis", "2023 Liu and Karydis", "2025 Bredenbeck et al.", "2026 REFLEX (average)"],
"series": [
{ "name": "Autonomous recovery speed", "values": [2.0, 2.58, 3.5, 3.7, 5.03] }
],
"caption": "Previously reported values are from the related-work section of the REFLEX paper. Aircraft and conditions differ across studies",
"source": "arXiv:2610.15867"
}However, the figures in the chart were not compared under the same conditions. The speeds from earlier studies are quoted from the related-work section of the REFLEX paper, and we did not check the original sources. The aircraft structure, mass, auxiliary means of position estimation, and criteria for judging recovery success also differ. REFLEX's experiments involved near head-on collisions with a flat rigid wall, and position estimation used indoor motion capture.
The REFLEX paper also reports recovery from even faster collisions. In exploratory flight tests, the aircraft collided at 6.02 m/s and 6.05 m/s and recovered both times. The paper says this exceeds the highest speed of prior work demonstrating autonomous recovery by 64%. However, there were only two successes at the 6 m/s level, and the paper itself notes that this result alone cannot be used to assess recovery reliability or the space required.
There is also research on drones that survived faster collisions. According to the REFLEX paper, an aircraft with an icosahedral tensegrity structure, announced by Zha et al. in 2024, withstood a wall impact at 7.8 m/s, but required landing and re-takeoff to resume flight. Also, "HoLoArm" by Pham, Loianno, Ho, et al., published in the March 2026 issue of IEEE Robotics and Automation Letters (Vol. 11, No. 3), uses flexible arms inspired by the nodus structure of dragonfly wings. According to the paper's abstract, it withstood impacts up to 7.6 m/s, recovered in 0.3–0.6 seconds depending on impact direction and strength, and flew stably while carrying a 540 g payload. HoLoArm also runs a control policy obtained by reinforcement learning on PX4. Because each study defines recovery success differently, however, a simple speed comparison cannot determine which is better. What directly supports REFLEX's performance advantage is the comparison with standard PX4 on the same aircraft under the same conditions.
Squeezing Through 180 mm Gaps and Carrying Payloads, but Autonomous Real-World Flight Remains a Challenge
FlexiQuad's flexible structure has uses beyond crash resistance. According to the paper, an aircraft with a nominal width of 254 mm succeeded in flying through a 180 mm gap. That means it passed through a space only about 70% of its width. The mechanism for shrinking the ring uses a 9 g servo motor and string; thrust efficiency dropped from 79% to 68% when shrunk diagonally and to 73% when shrunk into a C shape. Meanwhile, passing through by deforming passively on contact with an obstacle was about 10 times faster than actively shrinking with the servo.
It also demonstrated payload handling. In an experiment, it grasped and carried a 136 g first-aid kit in flight and released it at a designated position. ETH also says it lifted a load heavier than its own weight. The large open space in the center is also distinctive in terms of building in a mechanism to hold payloads.
According to ETH, the material cost of the airframe is under 15 Swiss francs, equivalent to under about 2,900 yen at the reference rate published by the European Central Bank on November 3. If the ring frame is opened and flattened, it can reportedly be stored in a medium-sized suitcase. However, these statements about material cost and storability come from ETH's announcement, are not in the first version of the paper, and have not been independently verified.
The applications ETH cites include inspecting ship ballast tanks, search and rescue inside collapsed buildings, and environmental measurement in dense forests. Girardi also envisions stretching a net across the central void to capture suspicious drones. None of these has been demonstrated so far. ETH itself explains that operating outside the lab will require the ability to reliably perceive the surroundings and navigate autonomously.
For REFLEX too, conditions remain to be verified before deployment in real environments. The paper targeted near head-on collisions with a flat wall; oblique impacts, glancing contact with obstacles, and collisions with irregular surfaces have not been tested. Position estimation methods other than motion capture, and a mechanism that decides when to start recovery control using only onboard sensors, are also untested. The original FlexiQuad paper exists only as a first version awaiting peer review, and REFLEX is still under submission to ICRA 2027, so test conditions and the interpretation of the numbers may be revised through future review.
Girardi explains that with a stiff drone, a collision after delayed obstacle detection ends the mission, whereas FlexiQuad can survive collisions and may therefore fly at speeds too dangerous for conventional aircraft. This research has demonstrated high-speed flight, shock absorption on impact, and autonomous post-collision recovery individually. However, whether these can work together as one system in real environments has not yet been confirmed. Going forward, handling diverse collision conditions, including oblique ones, and surrounding perception, autonomous navigation, and recovery decisions using only onboard sensors will be key to evaluating practicality.