RL Octocopter -- Karolina Dubiel

In Progress<br>▶ First flight video

Fault-Tolerant RL Octocopter

Want to follow along? I'd love it if you tuned in for the updates I'll post on

and

Loading model…

drag to rotate · scroll to zoom

Building a custom octocopter from scratch -- 0 prior hardware experience,<br>idea to flying drone in 2.5 weeks. Designed in Fusion 360,<br>CNC-milled from G10 fiberglass and carbon fiber, and assembled by hand.<br>The end goal: an RL-trained controller that can sustain flight through single,<br>dual, and quad motor failures in simulation, deployed zero-shot to hardware.

Phases

FAQ

Phase I --CAD a custom octocopter design, CNC cut<br>it, and assemble the frame, motors, and propellers.

Phase II --Wire up electronics and take flight as<br>regular FC-powered octocopter.

Phase III --Develop and train an RL policy capable<br>of supporting the octocopter through regular flight and dual-motor failures.

Phase IV --Complete the sim-to-real transition and<br>achieve RL-powered flight. Sustain flight after shutting off 2 motors randomly in field tests.

Build Log

Jump to oldest ↓<br>Collapse all

below, paste it right after this ║<br>║ comment (newest entry first), fill in the title, date, ║<br>║ and content. Delete any optional fields you don't need. ║<br>╠══════════════════════════════════════════════════════════╣

Title of this update

Mon D, YYYY

Your update text. Can be multiple sentences.

More text in a second paragraph, if needed.

Link text →

╚══════════════════════════════════════════════════════════╝<br>-->

Day 17: Researching research

Jun 15, 2026

I have no undergraduate research experience. I was at only at Georgia Tech for two years before taking<br>the gap year that I'm still currently on, so I felt that there was never a good time to join a lab or<br>really get involved with the research community on campus. Given that I've never felt that grad school<br>immediately after undergrad is in the cards for me, I wasn't too drawn to research as a whole until this<br>project.

Getting more passionate about RL changed that for me. There's SO much fascinating literature about RL.<br>Today's update is about:<br>1. Research in relation to this<br>project<br>2. Some of the most interesting and helpful papers within this domain that I've read

Research ↔ this project

It's my dream for this project to become a published paper. Especially because I have no prior<br>research experience, there's a lot of things that would have to happen first. Obviously, the drone would<br>have to fly with RL (I would want this to be a paper based on not just sim data). I'd likely have to<br>choose some<br>research question and probably collect some non-RL benchmarks.<br>From there, I think publishing<br>independently would be the most streamlined option, but it would be incredible to be able to join forces<br>with a PhD student or lab as a whole to put the paper together.<br>This whole world of academia is a<br>little intimidating for me solely due to my lack of exposure to it, so the concept of cold emailing labs<br>and students does scare me a little, but that's good. I think anyone who is able to build a drone from<br>scratch without ever even flying a drone is also someone who can cold email some people about wanting help<br>with a paper. Leads very much appreciated, though -- if you have experience in academia, where do you see<br>this realistically going? What do you think are my best paths forward?

-->

Day 25: Being a physicist, aka swinging my drone from the kitchen table

Jun 23, 2026

I modeled and printed a GPS mount, which was the final component I needed to lock down my total system<br>mass. Now that the drone is fully complete, I was able to collect the full system identification data on<br>the drone. I haven't been able to dedicate as much time as I would've wanted to this drone for the past<br>~1.5 weeks, but I hope to lock back in now that it's a software-only task for the next little bit.

Yaw bifilar pendulum setup<br>(I promise the strings are parallel IRL)

GPS and receiver mount

Here are the results and calculations of my system identification. They're not completely perfect or<br>scientific, but I'm hoping domain randomization can make up for any inaccuracies. I'll be using<br>motor/propulsion characterization information from the manufacturer and have also collected CoM info.

I = (T² · M · g · d²) / (8π² · L) → I = 0.36974 · d² · T²

M = 1.177 kg · L = 0.39552 m · d =<br>half wire sep. (m) · T = time for 20 osc ÷ 20 (s)

Measurements

Each trial timed over 20 oscillations (T = total / 20). Wire separations were<br>measured as full width and halved for d.

7 trials each; pink = worst 2<br>dropped before averaging.

Axis

T1

T2

T3

T4

T5

T6

T7

Avg×20 (s)<br>T (s)

Wire<br>sep. (mm)<br>d (m)

Roll<br>12.97

12.87

13.15<br>12.92<br>12.9<br>12.82<br>12.92<br>12.906<br>0.6453<br>332.592<br>0.166296

Pitch<br>12.77<br>12.64<br>12.82

12.63<br>12.75

12.84<br>12.82<br>12.76<br>0.638<br>332.592<br>0.166296

Yaw<br>13.84<br>13.52<br>13.52<br>13.47<br>13.52<br>13.4<br>13.65<br>13.536<br>0.6768<br>380.314<br>0.190157

Results

Axis

T (s)

d...