Playbooks · Coding tasks
How MonoRace flies a drone on one camera and an IMU
Written November 4, 2026
A small neural network flying a racing drone at 500 control commands a second, seeing through one camera, with no external tracking to lean on. That is the result a TU Delft team reports in the MonoRace paper — won in a real competition, on one indoor track, with known gates.
What changed
MonoRace is an onboard drone-racing approach from Delft University of Technology's Micro Air Vehicles Lab. The paper, published in Science Robotics in 2026, describes a system that uses a single monocular, rolling-shutter camera and an inertial measurement unit (IMU), and generalizes to a competition environment without any external motion tracking system. It won the 2025 Abu Dhabi Autonomous Drone Racing Competition, held on 11–12 April 2025 in a 100 by 30 meter indoor hall, beating all competing AI teams and three human world champion pilots in a direct knockout tournament. The fastest completion time was 16.56 seconds over two laps through eleven gates, at speeds up to 100 km/h. The 2025 race and the 2026 paper are two different events: the race tested the system, and the paper reports how it worked.
How the loop works
Three parts do the work. A neural network segments the racing gates in each camera image. State estimation combines that with a model of the drone to work out where the drone is and how it is moving, from onboard sensing only. And a single guidance-and-control network, trained in simulation with reinforcement learning, maps the estimated state directly to motor commands, replacing the usual inner-loop controllers. That control network has only 3 layers of 64 neurons and runs on the drone's 32-bit flight controller at 500 Hz. The competition drone also carried an NVIDIA Orin NX companion computer connected to the camera. An offline optimization step refines state-estimation parameters, including camera calibration, using the known geometry of the gates and purely onboard flight data.
Strengths
- Minimal sensing, high speed. One forward-facing camera plus an IMU carried the whole job — perception, navigation, and control — at up to 100 km/h, with no motion-capture system in the hall and none allowed.
- Robust to bad input. The state estimation handles images that are 50% corrupted, and copes with accelerometer saturation, which occurred regularly during high-g maneuvers on the track.
- Three of four events won. MonoRace won the Grand Challenge time trial, the AI-versus-human knockout, and the drag race. It placed third in the multi-drone race.
- A reliable controller alongside the fast ones. The team trained several controllers and reports a clear speed-versus-robustness trade-off. The most reliable network, M23, completed 43 flights with an 88.4% success rate, with most mishaps described as development- or hardware-related rather than control failures.
Weaknesses and limits
- One hall, one known track. The gates, their shapes and sizes, and the track layout were fixed and known in advance, and the offline optimization step uses that known geometry. Nothing in the paper shows the same setup working on an unknown course.
- Organizer-supplied hardware. The competition drone, track, and hall were provided by the organizers, not built by the researchers. Results on different airframes, cameras, or lighting are not established by this paper.
- No collision avoidance. The third-place multi-drone result is attributed in the paper to the absence of collision-avoidance capabilities. The system is built for single-drone racing, not for flying among other drones.
- Simulation-trained. The control network learned in simulation and crossed the reality gap on this track. That crossing is the achievement; it is not evidence that simulation training transfers to every drone or environment.
- A peak speed, not a cruising speed. 100 km/h is the fastest the drone reached on the competition track. It says nothing about sustained speed, endurance, or performance outdoors.
What to do with it
For builders of embodied agents, the usable lesson is narrow and concrete: a small direct-control network on minimal sensors can win a demanding, real-world task when the task is tightly scoped and the training covers the conditions the system will meet. Test on your own hardware and your own course before drawing wider conclusions. Keep a reliable controller in the set alongside the fastest one, and measure the speed-versus-robustness trade-off on your own flights rather than assuming the fastest network is the one to ship. And treat known-environment tricks — gate geometry, pre-mapped tracks — as part of the system, not as free performance: remove them, and the result may not survive.
Limits of this reading
- The evidence is one competition, one track, and one hardware set. It is not a benchmark across drones, tracks, or conditions.
- Success rates come from development and competition flights, including hardware-related mishaps. They describe this team's campaign, not a general reliability figure for the approach.
- The paper does not claim outdoor flight, unknown environments, or multi-drone operation. Any of those would need its own demonstration.
Related playbooks
- Claude API effort control: change gear mid-stream
- Qwen Code 0.28 runs one shared agent loop
- T3 Code: drive your agents from your phone
- Add tests to an untested module with an agent
Sources
Back to: Coding playbooks hub · Also: Fifty agents built a bureaucracy; four beat the game · Google put a TPU in orbit and got Gemma 4 to answer