MonoRace

MonoRace is an onboard perception and control system for autonomous racing drones, developed at the Micro Air Vehicle Laboratory (MAVLab) of Delft University of Technology (TU Delft) in the Netherlands. It flies a quadcopter using only one forward-facing rolling-shutter camera and an inertial measurement unit (IMU), with no GPS and no external motion-tracking system, and it hands flight control to a small neural network that sends commands straight to the four motors 500 times per second.^[1][2][3] In April 2025 the system won the A2RL x DCL Autonomous Drone Championship in Abu Dhabi, taking the AI Grand Challenge time trial, the AI drag race and a knockout tournament against human pilots that, according to its developers, included three former world champions.^[2][4][5] The research was published in Science Robotics on 14 October 2026 as "MonoRace: Robust monocular AI for champion-level drone racing".^[1]

The system has four main parts: a gate-segmentation network and corner detector that locate racing gates in the camera image, an extended Kalman filter that fuses those detections with IMU data and a model of the drone, an offline procedure that tunes state-estimation parameters (most importantly the camera's mounting angles) from the drone's own flight logs, and a guidance-and-control network (G&CNet) trained with reinforcement learning in simulation.^[2] The authors report a fastest two-lap completion time of 16.56 seconds and a top speed of 28.23 m/s (about 102 km/h) on the competition track, which they describe as, to their knowledge, the fastest fully onboard autonomous flight reported to date.^[2]

Background: autonomous drone racing

In autonomous drone racing a flying robot must pass through a fixed sequence of gates as fast as possible, relying on its own sensors and computers. The MonoRace authors list the main difficulties: aerodynamic effects that are hard to model at high speed (which also limits the accuracy of simulators used for learning), motion blur and large frame-to-frame displacement in the camera images, and the need for close-to-optimal control on small onboard processors.^[2]

The paper also summarizes how the field's competitions have progressed. At the autonomous drone racing contests held alongside the IEEE IROS conference from 2016 to 2019, top speeds rose from 0.6 m/s (Daejeon, 2016) to 2.5 m/s (Macau, 2019). In 2019 Lockheed Martin and the Drone Racing League ran the AlphaPilot / AI Robotic Racing (AIRR) series, four events on standardized hardware with a $1 million prize. TU Delft's MAVLab won that championship with an average speed of 6.8 m/s and a peak of 9.2 m/s, while the University of Zurich team that finished second reached up to 8 m/s on the same platform.^[2] Even then, the two best AlphaPilot teams took almost twice as long as a professional human pilot to complete the track.^[11]

According to its authors and the MonoRace team, the first autonomous system to beat champion-level human pilots in head-to-head races was Swift, from the University of Zurich's Robotics and Perception Group, described by Elia Kaufmann and colleagues in Nature in August 2023.^[2][11] Swift raced Alex Vanover (2019 Drone Racing League world champion), Thomas Bitmatta and Marvin Schaepper on a seven-gate track inside a 30 x 30 x 8 m volume, winning 5 of 9 races against Vanover, 4 of 7 against Bitmatta and 6 of 9 against Schaepper, and set the fastest race time of the event.^[11] The MonoRace authors argue that two goals remained after Swift: removing any dependence on external infrastructure such as motion-capture systems, and cutting the sensor suite down to a single forward-looking camera plus inertial sensors.^[2]

Development and team

MonoRace was developed at the MAVLab, part of TU Delft's Faculty of Aerospace Engineering.^[3] All nine authors list the Control & Operation department of TU Delft as their affiliation.^[1] The arXiv version marks Stavrow Bahnam and Robin Ferede as equal contributors, with Ferede as corresponding author; TU Delft's press release calls Bahnam the study's main author.^[2][3] Christophe De Wagter led the competition team, described by the organizers as MAVLab's "team principal", and Guido de Croon is a professor at TU Delft.^[3][4][5]

The control approach grew out of MAVLab's earlier collaboration with the Advanced Concepts Team of the European Space Agency (ESA), which had proposed guidance-and-control networks as a way to approximate optimal control with far less computation than conventional algorithms.^[4] According to De Wagter, the team had to redesign both the controller's training procedure and the way it learns the drone's dynamics from onboard sensor data.^[4] The preprint was posted to arXiv on 21 January 2026 under the title "MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI"; the peer-reviewed version appeared in Science Robotics volume 11, issue 119, on 14 October 2026.^[1][2]

The A2RL x DCL competition

The Abu Dhabi Autonomous Racing League (A2RL), part of the Advanced Technology Research Council (ATRC), organized the inaugural A2RL x DCL Autonomous Drone Championship together with the Drone Champions League (DCL) at ADNEC's Marina Hall in Abu Dhabi, with a $1 million prize pool.^[5][6] The finals ran on 11 and 12 April 2025, alongside the DCL Falcon Cup Finals for human pilots at the same venue.^[2][7] According to the organizers, 14 international teams from the UAE, the Netherlands, Austria, South Korea, the Czech Republic, Mexico, Turkey, China, Spain, Canada and the United States qualified for the finals week.^[5]

The rules were designed to force fully onboard perception. Every team flew the same organizer-built drone, fitted with an NVIDIA Jetson Orin NX module, a forward-facing camera and an IMU.^[5] The camera was a single rolling-shutter CMOS sensor, external aids were prohibited at every stage, the hall had no motion-tracking system, and teams were not allowed to modify or precisely measure the track.^[2] The organizers also chose the track layout and the hall, so MonoRace had to be adapted to hardware and a course its developers did not control.^[2][3]

The indoor hall measured 100 x 30 m. The track had 11 square gates, each with a 1.5 m inner opening, including two double gates and a split-S, in which the drone passes through an upper gate and then dives back through the gate beneath it in the opposite direction. A run consisted of two laps, timed from the first passage of gate 1 to the second passage of gate 11.^[2] The organizers described it as a 170-metre course.^[5]

The paper concentrates on the first two events. MonoRace placed third in the multi-drone race because it had no ability to detect and avoid other drones.^[2]

System design

Hardware

The competition drone, supplied by the organizers, has a custom carbon-fibre frame, a total mass of 966 g, 5.1-inch propellers and eight orange LED strips for spectator visibility.^[2] Live Science reported its dimensions as roughly 30 x 28 x 10 cm.^[12] The battery sits at the front and the Jetson Orin NX at the rear, an elongated mass distribution that made gyroscopic coupling terms significant in the team's dynamics model.^[2]

The camera is an Arducam IMX219 rolling-shutter sensor with a wide 155 x 115 degree field of view, and the IMU, built into the flight controller, delivers accelerometer readings at 1000 Hz and gyroscope readings at 2000 Hz.^[2] Computation is split across two processors. The GPU-equipped NVIDIA Jetson Orin NX handles image processing, gate detection and sensor fusion, and sends the state estimate to the flight controller, an STM32H743 ARM microcontroller running at 480 MHz. The flight controller runs a 500 Hz local state-estimation filter and the control network.^[2]

Perception: finding the gates

The camera driver was modified to capture full-field-of-view images of 820 x 616 pixels at 90 Hz. After undistortion, the pipeline processes only a 384 x 384 pixel region, 29.2% of the image, so that it can keep up with 90 Hz on the Jetson.^[2] This "adaptive cropping" step uses the current state estimate to predict where the gate corners should appear and selects the area containing the two closest visible gates. Gates seen at a very oblique angle, or whose projected centre falls outside the image, are skipped. If the predicted corners fit inside a 384 x 384 window the region is cropped at full resolution; otherwise the image is first downscaled. Distant gates are therefore analysed at native resolution and large, close gates are shrunk.^[2] On the team's three fastest successful competition flights, adaptive cropping detected 36% more gate corners than simply resizing the whole image and produced 25% more frames with enough points for a pose solution.^[2] De Croon has described this step as a "covert attention model".^[14]

Gate pixels are found by GateNet, a U-Net-style encoder-decoder for image segmentation that produces five output maps at increasing resolutions; only the highest-resolution map is used at run time.^[2] It was trained for 100 epochs on a mix of 3,500 synthetic and 500 real labelled images. The synthetic images are made by warping gate faces onto background photos of warehouses, parking lots and empty halls, and training adds colour shifts, noise, lens distortion, motion blur and rolling-shutter blur so that the network copes with the blurred, distorted frames seen at racing speed.^[2]

A gate-specific corner detector called QuAdGate then turns the segmentation masks into precise corner positions. It rotates each mask to align with gravity using the estimated attitude, runs OpenCV's Line Segment Detector, extends the detected edges and takes their intersections as corner candidates, so that rounded or incomplete masks do not shift the corners.^[2] Each candidate gets a four-value descriptor sampled from the mask around it and is matched to corners predicted from the prior state estimate. A RANSAC fit of a constrained 2D affine transform then rejects wrong matches.^[2]

The drone's pose is computed with a Perspective-n-Point (PnP) solution. Unlike Swift, which solved a separate PnP problem for each gate, MonoRace combines corners from up to two gates in one optimization. Gates at different depths help separate translation from rotation, and pooling corners lets PnP reach its minimum of four points more often.^[2] On the fastest successful flight, enough corners for a PnP solution were detected in 963 of 1,620 frames, and 263 of those (27%) used corners from more than one gate.^[2] The full PnP result is trusted only when the gate is 2 to 5 m away and at least six corners are detected; otherwise the filter's attitude estimate is combined with the PnP translation.^[2]

State estimation

An extended Kalman filter with a 16-element state (position, velocity, orientation quaternion, and accelerometer and gyroscope biases) uses IMU readings for its prediction step and PnP positions and orientations as measurements.^[2] Measurements are timestamped and corrected for pre-calibrated transport and processing delays: each image updates the state at its capture time, and the result is propagated forward with the IMU samples that arrived since. In software-in-the-loop simulation of the competition track, with simulated delays of 17 ms for images and 0.5 ms for the IMU, this cut the RMS trajectory error to 0.103 m, compared with 0.289 m for a baseline that applied each visual update on arrival.^[2] PnP measurements that deviate too far from the filter's prediction, given its covariance and the number of corners, are rejected.^[2]

The filter also handles accelerometer saturation. The reinforcement-learning controllers could command thrusts of up to 7 g, and combined with frame vibration and sensor noise this drove the accelerometers past their 16 g range, where they sometimes returned wildly wrong values and made the filter diverge, mostly during the split-S and a sharp, high-speed turn after gate 1 on the second lap.^[2] MonoRace runs the same quadcopter model used in its training simulator alongside the filter. When the low-pass-filtered measured and predicted accelerations differ by more than 22 m/s², the filter switches to the model's accelerations on all three axes and inflates its position and attitude uncertainty so that it leans more on the camera.^[2] For M17, the controller that won the Grand Challenge, 5 of 7 flights succeeded; both crashes were caused by IMU saturation. The paper reports that IMU-only prediction gave 2 successes and 2 crashes, while the 3 flights flown with the model-based correction all succeeded.^[2]

Calibration from flight data

Because the hall had no external positioning system, the team had no ground truth for tuning or debugging. Its answer was to treat the racing gates themselves as the reference.^[2] Using the state estimate and the map of gate positions, the software reprojects where the gate masks should appear in each camera frame and compares them with the masks GateNet actually detected, using intersection over union (IoU) averaged over one or more flights. A higher IoU means a better state estimate, so the metric can be used both to compare code changes and as an objective for optimizing any parameter in the pipeline.^[2] De Croon summarized the idea as minimizing the difference between where the vision network detects the gates and where the state estimate expects them.^[14]

The main use was estimating the camera's extrinsic calibration, the angles between the camera and the drone body. The camera sat on a flexible TPU mount, so these angles changed after impacts, modifications or repairs, and the calibration setup available at the competition could not measure them precisely.^[2] After each crash or repair the team flew the drone around the track with one of its safer controllers and then ran Bayesian optimization on the logged flight to recover the angles, chosen because it gives good results from few samples under the competition's time pressure.^[2] This happens offline on logged flight data: the procedure estimates the camera angles that the state estimator uses, without any external measuring equipment.^[2]

Control: the guidance-and-control network

Guidance and control are handled by one network, the G&CNet, which maps the estimated state directly to the four motor commands and replaces the usual cascade of trajectory planner, position controller and inner-loop attitude and rate controllers.^[2] The concept was first developed for real-time optimal spacecraft landing (Sánchez-Sánchez and Izzo, 2016) and was later adapted to fast quadrotor flight.^[2] Earlier G&CNets had relied on state estimates from external motion-capture systems; MonoRace runs one on purely onboard, vision-based estimates, which de Croon called the first test of these networks in a closed control loop with onboard perception.^[2][14]

The network has three layers of 64 neurons and takes as input the drone's position, velocity and Euler angles in the frame of the current target gate, its angular rates, the four motor speeds, and the relative position and heading of the next gate.^[2] It runs on the flight controller's microcontroller at 500 Hz with full throttle authority, and the motor commands can swing from 0 to 100% in 2 ms.^[2] Removing the inner-loop controller, the authors write, eliminates its unpredictable behaviour near saturation and allows faster flight.^[2]

Training used Proximal Policy Optimization (PPO) from the Stable-Baselines3 library in a custom Gym environment.^[2] The simulator is a deliberately simple quadcopter model, integrated at 0.01 s steps, that captures motor dynamics, thrust, drag and moments. It adds a quadratic drag term and dependencies on the propeller's advance ratio and angle of attack, effects that matter at higher speeds. The nominal parameters were identified by linear regression on high-speed flight data, and inertia differences were measured by throwing the drone, spinning and without propellers, and regressing on the gyroscope data.^[2] De Croon told Live Science that this "stripped-down, pure physics simulator" contained no rendered images: the control network sees only an estimated state, so vision did not need to be simulated, and the policy "learns to fly from scratch" over millions of trials.^[12]

To make the policy survive the move from simulation to the real drone (sim-to-real transfer), every model parameter was randomized at the start of each training episode, by between 15% and 55% of its nominal value depending on the parameter and policy.^[2] Episodes started either on the ground near the first gate or at a random point anywhere in the flight volume. The reward combined progress toward the next gate (capped at a maximum speed for the slower, safer policies), a bonus for passing a gate, and penalties for high angular rates, off-centre gate crossings, letting the next gate drift more than 60 degrees from the camera axis, abrupt motor-command changes and crashes.^[2] The motor-change penalty was added after an early policy learned an almost pure bang-bang strategy, switching motors between minimum and maximum, and overheated them.^[2] The trained networks were deployed with zero-shot transfer.^[2]

Results

AI vs Human knockout

The best-ranked AI teams were paired with the best-ranked human FPV pilots from the DCL Falcon Cup in a single-elimination bracket of quarter-finals, semi-finals and a final.^[2][5] The organizers timed the races with a VTx-based radio-frequency system, so their times differ slightly from the onboard-estimated times used elsewhere in the paper.^[2] The bracket below is transcribed from the organizer-recorded results reproduced in the paper; "no time" means the competitor has no recorded completion time.^[2]

MonoRace was the only AI entry to win a heat against a human; the other three AI teams recorded no time in their quarter-finals.^[2] In the final neither MonoRace nor the human pilot has a recorded completion time, and the paper and the organizers both name MAVLab the winner of the AI vs Human event.^[2][5] The paper identifies the semi-final opponent as Killian Rousseau.^[2] The team flew M17 in the quarter-final and switched to M16, its fastest network, for the semi-final and final; networks were named after their best real-world completion time rounded down to the second, so M16 had completed the track in about 16 seconds.^[2]

The pilots' standing has been described in different ways. The paper calls them "three human FPV world champions", and TU Delft's April 2025 announcement said the drone beat "three former DCL world champions".^[2][4] The organizers' press release spoke of the drone outpacing "a world-leading human pilot" in the AI vs Human Challenge, while Gulf News reported that it defeated "three top DCL champion pilots".^[5][9]

Controller performance

The team trained several networks during the competition, trading speed for reliability: more aggressive controllers flew faster laps, more conservative ones crashed less.^[2] The authors report that real-world completion times closely matched the average times in randomized simulation, but simulated crash rates were less predictive, because the simulator did not model the vision-based state-estimation errors that occasionally caused real crashes.^[2]

Source: figure 2 of the preprint.^[2] The fastest completion time, 16.56 s by M16, corresponds to an average speed of 20.02 m/s, while the overall top speed of 28.23 m/s was reached by the slower-lapping M18.^[2] M23, the most reliable network, completed 38 of 43 flights, and the authors attribute most of its failures to development or hardware problems.^[2] During M16's fastest lap the measured forces, moments and motor speeds largely stayed within, and at times exceeded, the 30% randomization bounds around the nominal model, yet the controller still completed the lap, which the authors credit to heavy randomization during training.^[2]

The organizers reported the Grand Challenge result as two laps (22 gates) of the 170-metre course in 17 seconds.^[5] For comparison, the team from KAIST in South Korea reports in its own paper that it finished third in the Grand Challenge with a top speed of 43.2 km/h, second in the drag race at over 59 km/h and second in the multi-drone race.^[17]

Robustness to sensor faults

Camera interference was a recurring problem. The organizers' drone originally used long MIPI camera cables that bent under high thrust and touched other components such as a noisy SSD, corrupting images mostly during hard vertical acceleration. Across 10 recorded flights between 8% and 75% of images were corrupted.^[2] Two of those flights crashed: one after a complete camera failure, and one in which 75% of frames were corrupted and the drone flew more than 25 m without a valid image, passing one gate before hitting the next.^[2] MonoRace completed both laps with 50% of frames corrupted. The paper credits layered safeguards: GateNet finds no gates in the corrupted regions, RANSAC removes bad corner matches, and the Kalman filter rejects measurements far from its prediction.^[2] The cabling fault was later fixed.^[2]

Comparison with Swift

The MonoRace authors frame their work against Swift, which they call the previous state of the art.^[2] They describe Swift as combining Intel RealSense stereo visual-inertial odometry, neural-network gate detection on a forward camera, and a learned residual observation model trained with motion-capture supervision, with a deep-RL policy that outputs collective thrust and body rates, the same low-level commands human pilots use.^[2] TU Delft adds that Swift's races took place in a flight lab where the researchers controlled conditions, hardware and track, whereas at A2RL the organizers designed and ran both.^[4]

The authors give Swift's figure "for context", noting that it was reported on a different, 30 x 30 x 8 m track, while their own 28.23 m/s came from a 76 x 18 x 5.4 m track with a monocular camera, a saturating IMU and no external aids at any stage.^[2] The Swift paper itself had flagged the gap MonoRace targeted. Its authors noted that Swift's perception assumed the environment looked as it had during training, that it was not trained to recover after a crash, and that addressing these limits could allow the approach to be used in competitions where access to the environment and the drone is limited.^[11]

Limitations

In their discussion the authors list three open problems:^[2]

• Decoupled vision and control. The control network assumes a perfect state estimate even though the estimate is always uncertain, and the separation forces extra reward shaping, such as the penalty for not looking at the next gate. They suggest end-to-end learning as a possible remedy.
• Dependence on gate shape. Corner detection and gate-relative localization both rely on the gates being rectangular, whereas human pilots adapt to new gate shapes without training. A network that maps images directly to relative gate poses could help, but refining it in the real world without ground truth is harder; the mask-reprojection idea might supply that ground truth.
• No detection of other drones. This did not matter in the one-on-one human races, but in the four-drone race MonoRace took off first, had to lap the slower drones, flew into another team's drone and finished third.

Elsewhere the paper notes that the training simulator does not model vision-based state-estimation errors, which made simulated crash rates a poor guide to real ones (M16, for example, succeeded in 94% of simulated runs but 1 of 4 real flights).^[2] The paper does not list it among the open problems, but the pipeline also relies on a known course: localization works by measuring the drone's pose relative to gates whose positions in the world are given in advance as a flight plan. The authors add that generalizing the approach away from drone-race-specific elements such as racing gates would open up many more real-world applications.^[2]

Applications and outlook

The authors argue that the guidance-and-control network is small enough to run comfortably on a microcontroller while approximating time-optimal flight, at a fraction of the computation needed by pipelines that plan trajectories and then track them.^[2] They suggest that drone-racing AI could one day help human pilots train and discover new moves, as chess engines do for chess players.^[2] The paper acknowledges that such flight speeds immediately suggest military applications, offensive as well as defensive, while stating that the authors hope and expect the main applications to be societal ones, for example adjusting the reward so drones fly at energy-efficient speeds and slow down when safety requires.^[2] In TU Delft's announcement, de Croon said the same approach could be tuned for efficiency or safety rather than speed, on machines "ranging from small vacuum cleaning robots to huge ship-to-shore cranes".^[3] Live Science described the project's broader significance as a demonstration of physical AI: a robot perceiving, deciding and acting at high speed with limited sensing and computing power.^[12] De Croon has framed it in similar terms as a contribution to robust, real-world physical AI, and the TU Delft release suggests the principles could carry over to vehicles, mobile robots and humanoid systems, part of the wider field of embodied AI.^[3][14]

Publication, data and code

According to de Croon, a photograph from the research was used on the cover of the issue of Science Robotics that carried the article.^[14]

Coverage

TU Delft's MAVLab announced the competition win on 16 April 2025, reporting a top speed of 95.8 km/h, and described it as the first time a drone had beaten human pilots in an international drone racing competition.^[4] That announcement dated the races to "Saturday, April 14, 2025"; the paper and the Drone Champions League's published schedule give 11 and 12 April 2025, with the Falcon Cup final on Saturday 12 April.^[2][4][7] The organizers' release the same day highlighted MAVLab's three event wins and TII Racing's win in the multi-drone race.^[5] Khaleej Times and Gulf News carried the organizers' account, and TU Delft's campus newspaper Delta reported that the team had beaten 13 other AI teams as well as professional human pilots.^[8][9][10]

The Science Robotics publication in October 2026 brought a second round of coverage, including a TU Delft press release, heise online and Live Science.^[3][12][13] These reports generally gave the speed as "up to 100 km/h", the figure used in the paper's abstract; Live Science gave the top speed as about 102 km/h, matching the 28.23 m/s in the paper.^[1][2][12] Live Science noted that some sequences in a video about the project are computer-generated; de Croon said these were intentional visual effects, and the magazine reported that the only fully simulated imagery is in the segment illustrating training and that the rest is real footage, in places overlaid with what the perception system detects.^[12]

Later A2RL drone events

The A2RL Drone Championship returned for a second season in Abu Dhabi in January 2026, with a $600,000 prize pool and the same restriction to one monocular camera and an IMU.^[18] MAVLab recorded a 12.832-second lap in the AI Speed Challenge, second to TII Racing's 12.032 seconds, and won the Multi-Drone Gold Race.^[18] In that season's Human vs AI final, South Korean FPV pilot Minchan Kim beat TII Racing's autonomous drone: the best-of-nine contest was tied at four wins each before the AI drone clipped a gate in the deciding heat and could not recover.^[18][19]

Continue reading

• DDroneA drone is an aircraft that flies without a human pilot on board, controlled either remotely by an operator or autonomously by onboard computers using artificial intelligence and sensors.Mentioned 59 times above
• RRoboticsRobotics is the interdisciplinary study and engineering of robots: embodied machines whose mechanisms, sensors, computation, and actuators are organized to perform tasks in the physical world.Mentioned 8 times above
• Reinforcement learningReinforcement learning (RL) is a branch of machine learning in which an agent learns to make decisions by taking actions in an environment to maximize a cumulative reward signal, discovering good behavior…Listed in See also
• SSim-to-real transferSim-to-real transfer (also written as sim2real) is the process of training a policy, controller, or perception model inside a physics simulator and deploying it on a physical robot in the real world without…Listed in See also
• SSensor fusionSensor fusion is the process of combining data from multiple sensors, often of different types, to produce information that is more accurate, complete, or reliable than any single sensor could provide on its…Listed in See also
• DDrone AIDrone AI refers to the application of artificial intelligence technologies to unmanned aerial vehicles (UAVs), commonly known as drones, enabling them to perform tasks autonomously or semi-autonomously.Listed in See also

More from See also

• Computer Vision
• Model Predictive Control

More to read

• Inertial Measurement Unit
• Neural Network
• Simulation (in AI and robotics)
• Kalman Filter

Topics:Computer VisionEmbodied AIReinforcement LearningRobotics

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Bahnam, S. A., Ferede, R., Blaha, T. M., Lang, A. E., Lucassen, E., Missinne, Q., Verraest, A. E. C., De Wagter, C., and de Croon, G. C. H. E. "MonoRace: Robust monocular AI for champion-level drone racing." *Science Robotics* 11 (119), eaee9250, 14 October 2026. doi.org/...scirobotics.aee9250
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73 ^74 ^75 ^76 ^77 ^78 ^79 ^80 ^81 ^82 ^83 ^84 ^85 ^86 ^87 ^88 ^89 ^90 ^91 ^92 ^93 ^94 ^95 ^96 ^97 ^98 ^99 ^100 ^101 ^102 ^103 ^104 ^105 ^106 ^107 ^108Bahnam, S. A., Ferede, R., Blaha, T. M., Lang, A. E., Lucassen, E., Missinne, Q., Verraest, A. E. C., De Wagter, C., and de Croon, G. C. H. E. "MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI." arXiv:2601.15222, 21 January 2026. arxiv.org/...2601.15222
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Delft University of Technology. "The science behind world's fastest autonomous racing drone." 16 October 2026. tudelft.nl/...rlds-fastest-autonomous-racing-drone
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8De Wagter, C. (MAVLab, TU Delft). "Autonomous Drone from TU Delft Defeats Human Champions in Historic Racing First." 16 April 2025. mavlab.tudelft.nl/dronerace-defeating-champions
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15A2RL / ATRC press release via Business Wire. "Artificial Intelligence Triumphs in World's Most Sophisticated Autonomous Drone Race in Abu Dhabi." ITWeb, 16 April 2025. itweb.co.za/...mYZRXv9gZ6ZMOgA8
6. ^Abu Dhabi Media Office. "A2RL X DCL Autonomous Drone Championship advances global expertise in AI drone technology." 17 April 2025. mediaoffice.abudhabi/...ise-in-ai-drone-technology
7. ^1 ^2 ^3Drone Champions League. "DCL Falcon Cup Finals 2025: High-Speed Drone Racing Comes to Abu Dhabi." 7 April 2025. dronechampionsleague.com/...n-cup-finals-abu-dhabi
8. ^Tesorero, A. "Watch: Human pilot, AI race their drones in Abu Dhabi; who won?" Khaleej Times, 17 April 2025. khaleejtimes.com/...abi-human-pilot-ai-drones-race
9. ^1 ^2Kumar, A. "Watch: AI drone beats human pilot in Abu Dhabi's $1 million prize pool autonomous race." Gulf News, 18 April 2025. gulfnews.com/...e-pool-autonomous-race-1.500097717
10. ^van der Veldt, M. "AI Drone from TU Delft outpaces human pilots at World Championship in Abu Dhabi." Delta (TU Delft), 16 April 2025. delta.tudelft.nl/...orld-championship-in-abu-dhabi
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Kaufmann, E., Bauersfeld, L., Loquercio, A., Müller, M., Koltun, V., and Scaramuzza, D. "Champion-level drone racing using deep reinforcement learning." *Nature* 620, 982-987, 30 August 2023. nature.com/...s41586-023-06419-4
12. ^1 ^2 ^3 ^4 ^5 ^6Bradley, A. "World's fastest AI-powered racing drone outflies the best human pilots." Live Science, 19 October 2026. livescience.com/...ies-the-best-human-drone-pilots
13. ^Bünte, O. "Autonomous racing drone beats former world champions on course." heise online, 15 October 2026. heise.de/...mer-world-champions-on-course-11489330
14. ^1 ^2 ^3 ^4 ^5de Croon, G. LinkedIn post on the *Science Robotics* publication of MonoRace, October 2026. linkedin.com/...e-ugcPost-7521259411806531585-jlwY
15. ^Bahnam, S., et al. "Data and Code underlying the publication: MonoRace: Robust Monocular AI for Champion-Level Drone Racing." 4TU.ResearchData, 25 September 2026. doi.org/...039abd82-094c-4c93-baa1-e7446f562e0e
16. ^TU Delft. "MonoRaceGate: Drone Racing Gate Dataset." GitHub repository, accessed 31 October 2026. github.com/...MonoRaceGate
17. ^Azhari, M. B., Han, D., You, J. I., Park, S., and Shim, D. H. "Drift-Corrected Monocular VIO and Perception-Aware Planning for Autonomous Drone Racing." arXiv:2512.20475, 23 December 2025. arxiv.org/...2512.20475
18. ^1 ^2 ^3Singh, I. "AI drones get faster, and almost beat world's best pilot." DroneDJ, 27 January 2026. dronedj.com/...a2rl-human-vs-drone-championship
19. ^Technology Innovation Institute. "Watch: Human edges AI in dramatic drone racing finale in Abu Dhabi." 26 January 2026. tii.ae/...i-dramatic-drone-racing-finale-abu-dhabi

Show all 19 references