MonoRace-Inspired Autonomous Racing Drone: Self-Flying Quadcopter with Vision-Based Gate Navigation

An autonomous racing drone is a quadcopter that flies a marked racecourse with no pilot on the sticks — it sees the gates, decides where to go, and commands its own motors. On October 14, 2026, TU Delft's Micro Air Vehicle Lab published MonoRace in Science Robotics: a racing drone navigating on only a single forward-facing camera and an inertial measurement unit, no GPS and no external tracking, its neural network commanding the motors directly 500 times per second. At the 2025 Abu Dhabi Autonomous Drone Racing Competition it beat every rival AI team and three former human world-champion pilots at about 100 km/h. This project is the honest student-scale translation: a ducted 2.5-3 inch cinewhoop flying a 6-gate indoor course autonomously, swapping MonoRace's neural controller for a classical pipeline the student can build and defend — ArUco markers on cardboard gates, OpenCV detection on a Pi Zero 2 W, optical-flow hover stability and PID velocity control at indoor-safe speeds.

MonoRace's own numbers set the reference: roughly 100 km/h top speed, motor commands at 500 Hz, 5.1-inch propellers, a Jetson Orin NX companion computer with a neural network also running on the flight controller, and a single camera plus IMU as the entire sensor suite. This build keeps the idea and rescales the engineering: a prop-guarded indoor cinewhoop instead of a 5-inch racer, ArUco tags instead of a learned gate detector, a ~30 Hz vision-plus-PID loop instead of a 500 Hz neural controller, and a 4-6 m/s speed cap (design target) instead of 100 km/h. What transfers is the real engineering — gate detection from a single camera, sensor fusion for position awareness without GPS, and closed-loop flight through a timed course — plus the race format itself: the dashboard's human-vs-AI mode lets the student fly the same six gates in FPV and compare lap times against the autonomous mode, which is exactly the Delft-vs-champions comparison at lab scale.

A lap runs like this. The drone arms on the start pad and the Pi Zero 2 W starts its vision loop: each camera frame is scanned for ArUco markers, and the nearest gate's tag gives a range-and-bearing measurement. The optical-flow sensor reports ground-relative velocity, fused with IMU data for stable hover without GPS. The navigation node picks the next gate in the 6-gate sequence, computes the bearing error to its tag, and the PID controller issues velocity commands over the serial/MAVLink link, steering the drone through the gate while the speed cap holds a lab-safe 4-6 m/s. Each gate pass is timestamped and logged to the race-control dashboard, which draws progress on the live track map and stops the lap timer at the start/finish gate. Human mode replays the identical course in manual FPV with the vision loop disengaged, and the best-lap board compares autonomous against human laps side by side.

1. The drone arms on the start pad; the Pi Zero 2 W starts the vision loop and the optical-flow sensor locks onto the floor texture for hover stability.
2. Each camera frame is scanned for ArUco markers (OpenCV); the nearest gate's tag yields a range-and-bearing fix at about 24 Hz (design target).
3. The navigation node selects the next gate in the 6-gate sequence and computes bearing error to its tag; the PID controller converts that error into velocity commands.
4. Commands travel over the serial/MAVLink link to the flight controller, which steers the drone through the gate while the 4-6 m/s speed cap (design target) holds indoor-safe pace.
5. The PMW3901 optical-flow sensor plus IMU fusion tracks relative motion between gates, where no GPS exists and the camera sees only the next tag.
6. Every gate pass is timestamped and pushed to the race-control dashboard, which draws live progress on the track map and stops the lap timer at the start/finish gate.
7. Human mode replays the identical course in manual FPV with the vision loop disengaged; the dashboard's best-lap board compares autonomous vs human laps.

Tech stack:

• 2.5-3" ducted cinewhoop frame + flight controller (prop-guarded indoor racer)
• Raspberry Pi Zero 2 W (companion computer · OpenCV gate detection)
• Forward-facing camera module (Pi camera class)
• PMW3901 optical-flow sensor + flight-controller IMU fusion
• OpenCV ArUco fiducial detection + PID velocity control
• MAVLink/serial offboard command link
• Race-control web dashboard (lap timer, track map, gate-pass log)
• 6-gate ArUco-tagged cardboard course kit

The drone's capabilities are four. First, autonomous gate navigation: the forward camera feeds OpenCV ArUco detection on the Raspberry Pi Zero 2 W at about 24 Hz (design target), each gate's fiducial marker giving the drone a range-and-bearing fix that the PID controller turns into velocity commands — the student-scale stand-in for MonoRace's neural gate detector. Second, GPS-denied position awareness: a PMW3901 optical-flow sensor fused with the flight controller's IMU holds hover and tracks relative motion indoors where GPS cannot reach, the same no-GPS constraint MonoRace solved with learned estimation. Third, the human-vs-AI race mode: the race-control dashboard times every lap in both autonomous and manual FPV modes on the identical 6-gate course, logging every gate pass with timestamps so the student can present a genuine human-against-AI comparison. Fourth, indoor-safe racing engineering: full prop guards, a netted-course operating procedure, a 4-6 m/s speed cap and an RC-link kill switch, so the whole build is lab-legal from day one.

• [Autonomous gate navigation] A forward-facing camera feeds ArUco fiducial detection (OpenCV) on the Raspberry Pi Zero 2 W at about 24 Hz (design target); each gate's tag gives a range-and-bearing fix that the PID controller converts into velocity commands, flying the 6-gate course with no pilot input.
• [GPS-denied position awareness] A PMW3901 optical-flow sensor fused with the flight controller IMU holds position and tracks relative motion indoors, where GPS is unavailable — the same no-GPS constraint MonoRace solved, answered here with classical sensor fusion instead of a learned estimator.
• [Human-vs-AI race mode] The race-control dashboard times every lap in autonomous and manual FPV modes on the identical 6-gate course and logs every gate pass with timestamps, producing a genuine human-against-AI lap comparison for the report and viva.
• [Indoor-safe racing engineering] Full prop guards, netted-course procedure, a 4-6 m/s speed cap (design target) and an RC-link kill switch keep every test flight lab-legal; nothing in the build assumes outdoor or unsupervised flight.
• [Race-control web dashboard] Served by the companion computer: live track map, lap timer, best-lap board, gate-pass event log and telemetry (speed, battery, detector rate) in one view.
• [6-gate ArUco course kit] Six cardboard racing-gate hoops with printed ArUco fiducial markers and floor layout plan, so the course is reproducible in any lab or classroom.
• [Tunable vision-PID pipeline] Gate-detection thresholds, PID gains and speed cap exposed as a single config file, with a tuning procedure the student runs and documents — the tuning log is original data for the report.

• Assembled racing quadcopter: ducted cinewhoop frame with prop guards, flight controller, forward camera, PMW3901 optical-flow sensor, Raspberry Pi Zero 2 W companion computer
• Six ArUco-tagged cardboard racing gates with floor layout plan
• Complete software: OpenCV gate-detection pipeline, PID navigation node, MAVLink offboard bridge, race-control web dashboard (working demo UI included)
• Wiring diagram and airframe assembly layout
• Tuning procedure for detector thresholds, PID gains and speed cap
• Component list with Indian supplier sourcing notes for the frame, Pi Zero 2 W, camera and flow sensor
• Project report PDF (vision navigation, sensor fusion, PID tuning, human-vs-AI race results)
• PPT presentation for final review
• Viva Q&A preparation document (ArUco vs learned detection, optical flow, PID control, drone regulations)