A drone race in Abu Dhabi last year ended with a result that will bother a lot of pilots: the winner was not a person. MonoRace, an onboard system built by a team led by Stavrow A. Bahnam and Guido C. H. E. de Croon, took the Abu Dhabi Autonomous Drone Racing Competition 2025 by beating every competing autonomous team and three world champion human pilots in a direct knockout tournament. The result was published in Science Robotics on October 14, 2026.
The headline number is 100 kilometers per hour. That is the top speed MonoRace reached on the competition track, and it did so with a sensor suite that would make most robotics labs nervous: a single monocular rolling-shutter camera and an inertial measurement unit. No stereo rig. No external motion tracking. No room full of cameras telling the drone where it was.
That last point is the part that matters for the field. Autonomous drone racing has long relied on external tracking systems in controlled indoor spaces, which is a polite way of saying the robots were cheating with help from the building. MonoRace generalizes to a competition environment without any of that. It flies on its own data, gathered by its own sensors, in real time.
How it actually works
The system has two halves. The first is state estimation. MonoRace combines a neural network that segments the racing gates with a drone model, so the aircraft can figure out where it is by recognizing the gates ahead of it rather than by triangulating from fixed cameras. An offline optimization step then refines the state estimation parameters using the known geometry of those gates, working purely from onboard flight data. That step is what tightens up the extrinsic camera calibration, the small misalignment between where the camera thinks it is pointing and where it actually points.
The second half is control, and this is where the design gets opinionated. Instead of the usual stack of inner-loop controllers handing commands up a chain, MonoRace runs a single neural network that sends motor commands directly. It runs at 500 hertz on the flight controller. Five hundred decisions per second, each one a direct output to the motors, is fast enough that the aircraft can correct itself mid-maneuver rather than waiting for a controller hierarchy to catch up.
It is not clear this will hold at higher speeds or on tracks the network has never seen. The paper is honest about that. But the Abu Dhabi result is a data point, not a demo reel, because it happened in a knockout tournament against pilots who have won world championships.
What the Abu Dhabi run looked like
The competition track had the usual racing problem set: gates to pass through at speed, and the aerodynamic mess that comes with high velocity. At 100 km/h, airflow around a small quadcopter stops being a clean physics problem and starts being a guess. Complex aerodynamic effects become hard to model, which undermines both conventional control and the simulators used to train learning-based methods.
Vision suffers too. At racing speeds the camera sees large interframe displacements, high optical flow, and motion blur. A rolling-shutter sensor makes this worse because it captures the image line by line, so a fast-moving scene warps. MonoRace coped with camera interference and IMU saturation during the event, two failure modes that would have ended a lesser system's run.
The human pilots were not amateurs. Three of them were world champions, and they lost in direct knockout matchups. That framing matters. This was not a benchmark on a test track with no pressure. It was a tournament where a bad gate meant elimination.
Why this is more than a racing stunt
Drone racing is a proving ground, and the researchers say so plainly. The same problems that make racing hard, high-speed perception, state estimation, planning, and control under tight compute and time budgets, are the problems that matter for any drone that has to move fast in the real world. Delivery drones dodging obstacles in a city, inspection drones in a wind farm, search-and-rescue aircraft in smoke. None of them get the luxury of a motion-capture room.
The monocular constraint is the part worth watching. Stereo cameras add weight, cost, and compute. A single rolling-shutter camera is cheap, light, and already on most drones. If MonoRace's approach transfers, it lowers the hardware bar for autonomous flight rather than raising it.
There is a catch, and it is the honest kind. The offline optimization step uses known gate geometry. In a race, the gates are fixed and measured. In a delivery route or a disaster zone, the geometry is not handed to you. The researchers' next problem is whether the same calibration trick works when the environment is not pre-measured.
What comes next
The paper is out, the competition is won, and the obvious question is deployment. The Abu Dhabi Autonomous Drone Racing Competition 2025 is over, but the method is now public and the hardware requirement is a camera and an IMU, which means replication is cheap. Other teams will try to beat it, and some will. That is how this field moves.
For the rest of us, the useful signal is simpler. A single-camera drone just beat three human world champions at 100 km/h. The hardware was not exotic. The compute was onboard. The system did not need the room to help it. That is the bar for autonomous flight, and it just moved.
For the full technical details, see the original paper in Science Robotics. The competition results and team list are documented by the Abu Dhabi Autonomous Drone Racing Competition organizers.
Readers tracking this space can follow more coverage in our robotics and drones category.
The next test will not be on a race track. It will be in a place with no gates, no measured geometry, and no second chances. That is where MonoRace either proves it is more than a racing trick or quietly fails. Either outcome teaches the field something. It also gives the rest of the industry a useful number to beat: 100 km/h, one camera, no external help.