MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI

TL;DR

The paper tackles fully onboard monocular autonomous drone racing without external tracking, introducing MonoRace, a compact perception-control pipeline where a small Guidance-and-Control Network (G&CNet) directly outputs motor commands from onboard state estimates. The system fuses GateNet-based monocular gate segmentation with QuAdGate corner localization and an EKF that incorporates a dynamic model to handle IMU saturation, enabling robust, low-latency control at 500 Hz. It pairs this perception with a reinforcement-learning controller trained in a high-fidelity quadrotor simulator, using extensive domain randomization and an IoU-driven offline calibration mechanism to refine extrinsic camera parameters from flight data. The approach demonstrated state-of-the-art performance at the 2025 A2RL Grand Challenge, achieving speeds up to 100 km/h, surviving substantial image corruption and IMU saturation, and beating champion-level human pilots in direct knockout heats. Together, these contributions show that a lean monocular onboard system with a compact neural controller can achieve real-time, high-speed autonomous flight with broad applicability beyond drone racing.

Abstract

Autonomous drone racing represents a major frontier in robotics research. It requires an Artificial Intelligence (AI) that can run on board light-weight flying robots under tight resource and time constraints, while pushing the physical system to its limits. The state of the art in this area consists of a system with a stereo camera and an inertial measurement unit (IMU) that beat human drone racing champions in a controlled indoor environment. Here, we present MonoRace: an onboard drone racing approach that uses a monocular, rolling-shutter camera and IMU that generalizes to a competition environment without any external motion tracking system. The approach features robust state estimation that combines neural-network-based gate segmentation with a drone model. Moreover, it includes an offline optimization procedure that leverages the known geometry of gates to refine any state estimation parameter. This offline optimization is based purely on onboard flight data and is important for fine-tuning the vital external camera calibration parameters. Furthermore, the guidance and control are performed by a neural network that foregoes inner loop controllers by directly sending motor commands. This small network runs on the flight controller at 500Hz. The proposed approach won the 2025 Abu Dhabi Autonomous Drone Racing Competition (A2RL), outperforming all competing AI teams and three human world champion pilots in a direct knockout tournament. It set a new milestone in autonomous drone racing research, reaching speeds up to 100 km/h on the competition track and successfully coping with problems such as camera interference and IMU saturation.

Figures (7)

• Figure 1: (A) Close-up of the supplied externally made competition robot with its 5.1-inch propellers, its drone-race autopilot board with normal IMU, a single CMOS rolling-shutter camera connected to a GPU-equipped NVIDIA ORIN NX companion computer, and 8 bright orange LED strips to provide good visibility for the spectators. The robot is looking at an 'A2RL x DCL' competition gate. (B) A time-lapse of the robot performing the split-S maneuver. (C) Overview of the entire track. The races consisted of two laps through eleven gates. Four robots are shown at the four start positions during a multi-robot race. (D) Results of the AI against human races.
• Figure 2: (A) G&CNet completion times: real flights versus simulation. (B) Fastest lap trajectory top view (M16) and the corresponding measured forces, moments, and actuator outputs versus the nominal model.
• Figure 3: Self-supervised refinement from onboard data only using mask-based extrinsic optimization. (A) Camera extrinsics optimization in the Software In the Loop (SIL) simulator. The estimated extrinsics align within $\approx 0.5^\circ$ of ground truth in only 40 steps optimizing a single log. (B) Result of the self-supervised refinement applied to camera extrinsics optimization on real flight data for only $40$ steps.
• Figure 4: Onboard sensor data corruption. (A) Comparison of the accelerometer-based versus model-based predicted trajectories, thrust and gate reprojections during a split-S with accelerometer saturation. (B) Successful dual lap completion despite 50% image corruption (left) and single lap completion with 75% image corruption from camera interference-caused purple horizontal bars, outdated top pixels and vertically shifted parts of the image.
• Figure 5: Overview of the pipeline and physical location of various elements. An IMX219 camera image is grabbed and timestamped on the Orin. Adaptive state-based cropping selects the part of the wide-field-of-view image with the best statistics for detecting corners. Gate segmentation is performed on the GPU, and subsequently, precise corner detection is performed from edge identification. After outlier rejection, the relative pose is computed with respect to the assumed gate location. After sensor fusion with delay compensation, the state is sent to the flight controller, which runs a 500Hz local state estimation filter and the direct-to-motor neural G&CNet, which was trained in RL simulation with randomization and is being executed with zero-shot transfer.