Autonomous Perception
In researchAIDrone
Training multirotors to fly racing gates from onboard cameras and sensors alone, with no external motion-capture room.
Research & Notes
Ongoing research directions, open engineering tools, and field notes on onboard visual autonomy, endurance physics, and human-interpretable flight software.
Active work
Autonomous Perception
In researchTraining multirotors to fly racing gates from onboard cameras and sensors alone, with no external motion-capture room.
Human Visibility
LiveReconstructs 3D flight lines and attitude live in the browser from stick inputs and telemetry so people at the field can follow intent.
Endurance Physics
2:25:02 verifiedStrict mass and watt accounting on the opposite extreme from sprint racing, staying airborne for 2 hours, 25 minutes and 2 seconds in its first public outing.
Training Tools
App Store · RebuildingPractical tools from the start line: randomized start timer on iOS and Apple Watch, alongside a strength-of-field pilot rating model.
Researchers at the University of Zurich had 21 pilots race in a simulator while wearing eye trackers. About a second and a half before the drone reached a gate, the pilot's eyes were already on it, and in turns they looked at the inside of the line. In 2024 the same lab flew a quadrotor through gates from camera images alone, with no inertial sensor, at up to 40 km/h.
This is the question I am trying to answer, and I have not answered it. AIDrone, the project I am working on, starts here: a quad that flies a racing track by itself using only its onboard camera and sensors. It is still research.
A racing quad usually empties a battery in two and a half minutes flat out. At the end of 2023 in Fuzhou, we entered the Qiaozhihui endurance challenge, which asks the opposite question: how long can one aircraft stay in the air without landing? It was the first public outing for our long-endurance battery system, and we stayed up for 2 hours, 25 minutes and 2 seconds to win it.
In a sprint race your eyes lock onto the next gate and your thumbs move a dozen times a second. Over two and a half hours you stand at the edge of the field watching voltage, current draw and wind, knowing a single mistake in 145 minutes ends the run. Fast and slow look like different sports, but the ledger is the same: where every gram and every watt is going.
When I joined Super Pulse, drone racing in China did not have an official circuit yet. The crew was a mix of everyone who wanted to fly. We pooled orders to cut carbon frames, crashed them, soldered them back together at the field, and put them back on the start line.
Later, on the national team, the long table in the tent had students, engineers, factory owners, sponsors and twelve-year-old rookies sitting shoulder to shoulder. Talking with the people who run factories and brands taught me why a part costs what it costs and what it takes to ship a batch; watching the youngest pilots taught me that they are not afraid of lines older pilots overthink.
Chasing 0.1 seconds on a lap also meant nights spent with Betaflight contributors from different countries, working through filtering and PID code in broken English full of jargon. We misunderstood each other plenty of times, laughed it off, flashed the firmware and flew a pack. Once the quad was in the air, everyone understood the same thing.
What people call talent is mostly one person who stays with a thing they love and keeps putting in the work. That is how I got here.
When a lap will not get faster, I do not grind it out. I stop, look at the line across the whole track, then go back to the tune and the filters, fix what is wrong and fly it flat out. I tune my own quads and change the code myself, so I know in the air what is off. Races are often decided by 0.1 seconds.
I used to want first place in everything, and ended up not knowing what I actually liked. So I do not ask my pilots to be first at everything. If it does not work, we go again. What I want them to learn is to keep going at the thing they love, and to live a full life doing it.
The race starter, the LED frame and the rankings all began as something I needed and could not find. The first version is usually rough. The people using it tell me what is wrong, and I keep fixing it.
When my old high school asked me in 2025 to look back on applying to college, I remembered a comparison one of my teachers once made: choosing a university is like choosing a partner — it has to fit both sides, and you cannot hire a matchmaker to do the dating for you.
In high school I skipped the usual test-prep and summer-program race. When I noticed how scattered our student organization was, I spent a semester interviewing students across the school, ran a one-year pilot in my own homeroom with my teacher's backing, and brought the data to the principal, who adopted it school-wide. The rest of my hours went into flying, hardware and code. That spring I was offered a full scholarship to the Rotman Commerce program at the University of Toronto and admission to Computer Science at the University of Michigan, before choosing Physics and Computer Science at USC.
Physics and code teach me how to work through a problem at the microscopic level, one step and one signal at a time; economics and business teach me how to read the broader direction. Looking back, it matters much less how many boxes you check on someone else's form than finding the one thing you are willing to stay with for ten years.