Development and Evaluation of a Quadcopter Safety Mechanism for the Event of a Single Motor Failure

Over the course of about 9 months, I designed, built, and tested a safety mechanism for quadcopters.

Introduction

Quadcopters have a design flaw nobody really talks about, and it’s what got this whole project started. Over 855,000 drones are registered with the FAA in the US alone (HireDronePilot, 2026), and they’re everywhere because they’re cheap and they work well, making up about 71% of the global drone market (Fact.MR, 2023). All four motors have to work for the thing to stay balanced. The second one dies, the drone can’t compensate, and it spins out and crashes, sometimes at speeds over 80 mph. Parachutes and airbags are the main safety options right now, but they typically need 1–2 seconds just to deploy and inflate (Indemnis, 2022; ParaZero, 2023), and a drone falling from a normal height only has about 2–4 seconds before it hits the ground, so that’s usually not enough time for either to work. UAVs also fail at a rate roughly 100 times higher than manned aircraft, with motor failures being one of the leading causes (Ciani et al., 2018). So the question I kept coming back to was: could I make a quadcopter fall slower, and for longer, so that if a motor does fail, there’s a window of time for something else to happen before it hits the ground.

Starting the Research

I looked into a few different approaches before landing on thrust vectoring. The reason a single motor failure is so hard to recover from is that the imbalance it creates doesn’t self-correct on its own, a small roll leads to more asymmetric thrust, which leads to more roll, and it cascades until the drone crashes, without some kind of active fix (Gibiansky, 2012). Thrust vectoring addresses that directly: angling the motor opposite the one that failed redirects its thrust sideways instead of straight down, which fights back against the spin causing the cascade in the first place. Zaludin (2024) had already shown in simulation that a 45-degree tilt on the opposing rotor could reduce descent rate, and other tilted-rotor research backed up that redirecting a motor’s thrust can restore control after a failure (Ryll et al., 2015). But nobody had built it and tested it in real life, which is what got me interested.

Thrust Vectoring Mechanism Design

Building the Drone

Getting the drone to fly at all took way longer than I thought it would. I used a HAWKS WORK F450 kit so I wasn’t also designing an airframe from scratch on top of everything else. But partway through, I realized Lua, the scripting language I needed, wasn’t supported on the flight controller I started with, the Pixhawk 2.4.8, so I had to switch to a Pixhawk 6X. That meant redoing pretty much all the wiring and configuration from the ground up, and it caused its own problems, like the motor wiring order getting flipped somewhere in the process, which I didn’t catch until several failed test flights later where the drone just kept flying wrong and I had no idea why.

Practicing Flying the Drone

Making the Code

The code itself also posed a lot of issues, Lua as a language isn’t widely used, so a lot of it was based off old forums and threads, as well as their documentation. I went through dozens of different attempts and revisions before landing on something that worked, most of them small tweaks that failed for reasons I didn’t figure out until later. A few stand out. Early on I tried redefining one of the motors as a servo through the flight controller’s parameter system. That made it controllable, but it also pulled the motor out of the stabilization loop completely, so the flight controller had no way to compensate for it being gone, and stable flight was impossible. Another time I tried setting up an “emergency kill” switch to just cut the motor, but no matter what I changed, it wouldn’t work without killing all four motors at once, which defeated the point. What finally worked, after enough failed versions, was writing a Lua script that runs on a loop, watches an RC channel, and when that channel crosses a certain value, kills the failed motor and moves the servo to the vectoring angle at the same time, without ever removing that motor from the flight controller’s awareness. The other three motors keep actively stabilizing the whole time.

At first, I had to manually flip the switch back to turn the motor back on once the drone had fallen far enough, which meant timing it myself every single trial and hoping I didn’t mess it up. Eventually I coded it so the drone would reactivate the motor on its own once it dropped to 30 meters, using its barometric altitude reading to trigger it automatically. That one change made testing far less stressful (once it started working), because it guaranteed that the drone would land safely.

Developing the Mechanism

Getting the code right was still only a step , because the script was useless without something physical for the servo to actually move. The mechanism itself went through many prototypes before it was usable. The first version cut straight into the arm and used a hinge, held together with a servo and a piece of wire, so the arm would fold at the failure angle. It didn’t work, the servo was too slow and cutting into the arm wrecked its structural integrity.

So I scrapped that and switched to a U-shaped bracket that the motor sits inside, resting on top of the arm instead of cutting into it, with the servo rotating the whole bracket to the target angle. Even that took a lot of tries to get right, the motor’s bottom plate kept rubbing against the inside of the bracket and generating heat, so I drilled a hole in the bottom to let it spin freely. The tolerance had to be exact too, tight enough that the motor wouldn’t wobble, loose enough that the servo could still rotate it. It took more than five printed versions before I got one that would work.

Testing

All the testing happened at a field near me called 60 Acres, and it wasn’t one or two days of data collection, but rather it was spread out over the whole project, because I was testing different code versions and mechanism tweaks as I went.

Image of the Testing Setup

I’d fly the drone up to 70m, kill a motor mid-air on purpose, have the mechanism move to the desired angle, and watch it fall until it hit the safety threshold of 30m. I ran a control condition too, motor cut with no thrust vectoring at all, just to have a real baseline to compare everything else against. The first round of real testing was on the 30-degree and 45-degree tilts.

Mission Planner Software used

Trials got cut short more than once. I’ve had the drone go to the 70 meters, and then all 4 motors turned off and fell due to a coding error. While quite the messy process, I was able to resolve these issues time and time again to get my data for both angles.

CSRSEF

This project was also going to be submitted for the Central Sound Regional Science and Engineering Fair (CSRSEF), and the Washington State Science and Engineering Fair(WSSEF). This was my first time doing anything like a science fair, so going in I didn’t know what to expect. I was still collecting and analyzing data right up until the last minute, spending my lunches at school graphing everything out, it was a grind right up to the wire, to the point I barely had time to practice a script for presenting it.

Project Setup for CSRSEF

But when I got to CSRSEF it turned out to be a lot less stressful than I’d built it up to be in my head. It was talking about something I’d spent months on with someone who wanted to know more about it, and I was grateful they saw value in what I did. I ended up placing first in my category, Applied Physics and Mathematical Sciences, and finished in the top 15 out of over 400 projects.

Testing (Round Two)

At that point I still hadn’t finished testing dynamic vectoring, the version that adjusts itself in real time instead of sticking to a fixed angle, so that became the priority for the next phase. I went back out to 60 Acres and kept testing, running more trials on the 30-degree and 45-degree conditions alongside the new dynamic one. More parts broke. A few trials got scrapped entirely because the drone fell too fast or something on the mechanism failed mid-fall.

Broken Drone Following a Failure

But the trend held across all three versions, every one of them made the drone fall slower and stay in the air longer than the control did, by a wide enough margin that it wasn’t luck. I used Cohen’s d test to check the effect size, and it came out large across almost every condition (McKnight & Najab, 2010). Even the control trials fell slower than pure free fall would predict, which tracks, as the drag from the frame and propellers resists the descent even with the motors off (NASA Glenn Research Center, 2021). I still want a much bigger sample size before I’d call this a finished safety system, and there’s more tuning to do. But overall, the mechanism does work, as the quadcopter that would otherwise spin and slam straight into the ground can be made to fall slower, with less force behind the landing.

WSSEF

Going to WSSEF meant more field days, more broken parts, more re-flying the same failure over and over in whatever weather I could get, on top of finishing up the dynamic vectoring testing. I stood at my poster for hours, walking judges and other students through the same story every time, the flight controller swap, the failed code attempts, the prototypes that didn’t work. Some judges knew drones well and wanted to get into the PD controller and the yaw correction, and some just wanted to know more about my project.

Project Set Up at WSSEF (Photo Creds: WSSEF)

Ultimately, my project ended up placing at both the Central Sound Regional Science and Engineering Fair and the Washington State Science and Engineering Fair. Here’s what it won:

• 1st Place, Applied Physics and Mathematical Sciences (Category Winner) — Central Sound Regional Science and Engineering Fair
• CSRSEF Finalist (Top 15 out of 400+ projects) — Central Sound Regional Science and Engineering Fair
• 1st Place Award — Washington State Science and Engineering Fair
• Excellence in Aeronautics Award — American Institute of Aeronautics & Astronautics (AIAA Pacific Northwest)
• Launch Scholarship ($2,000/year, renewable up to $8,000 total) — Embry-Riddle Aeronautical University
• “Sights of Flight” Award — Museum of Flight
• STEM Award — Navy League of the U.S. (Bremerton Olympic Peninsula)
• Naval Science Award — Office of Naval Research

Photo Creds: WSSEF/CSRSEF

Conclusion

Looking back, I’m just really grateful for this whole experience. I learned so much and gained a lot of different skills. None of it would’ve happened without my parents support and them being there for every single field day, rain or shine. I’m also so grateful to all my teachers who supported me on this crazy journey. There’s still so much more testing and tuning ahead, and I can’t wait to see where this goes next!

References:

Ciani, L., et al. (2018). Reliability and maintenance analysis of unmanned aerial vehicles. Sensors. Fact.MR. (2023). Quadcopter market share, size & forecast. Gibiansky, A. (2012). Quadcopter dynamics and simulation. HireDronePilot. (2026). Drone statistics 2026: Expanded edition.Indemnis. (2022). Nexus parachute recovery system.

McKnight, P. E., & Najab, J. (2010). Mann-Whitney U test. In The Corsini Encyclopedia of Psychology.

NASA Glenn Research Center. (2021). The drag equation.

ParaZero. (2023). SafeAir parachute system overview.

Ryll, M., Bulthoff, H. H., & Giordano, P. R. (2015). A novel overactuated quadrotor UAV. IEEE Transactions on Control Systems Technology.

Zaludin, Z. A. (2024). Fault tolerance conceptual strategy for a quadcopter drone with rotor failure. Asian Review of Mechanical Engineering.