2025 · Hardware / Robotics
Drone Litter Mapping + Recovery System
An F450 drone that finds litter from the air, GPS-tags it, posts it to a mapping backend — while a separate walking robot handles the recovery.
Published Paper · YOLOv5 · 90% Detection
The problem
Nobody has an accurate map of where litter actually accumulates. Cleanup crews get allocated by intuition — park X must be bad — and by the time anyone walks it, points a GPS at the piles, and gets back to the office to make a list, the effort has cost more hours than the cleanup. And on the collection side, a drone can't safely land next to litter: rotor wash scatters it.
The constraint
Two very different problems. From the air, a soda can at 10–15 m is a few pixels among grass, gravel, and shadows, and a detection is useless without an accurate real-world coordinate to attach to it. On the ground, the recovery hardware had to walk (not roll) and stay stable — the walking robot's paper-measured speed was ~3 cm/s, which is safe but slow, and a bad gait resets a battery-hungry platform.
What I built
An F450 quadcopter: Pixhawk 2.4.8 with ArduPilot runs 400 Hz-class PID stabilization while a Raspberry Pi 4 companion computer connects over a hardware UART (57600 baud, via /dev/ttyS0). On the Pi, DroneKit/PyMAVLink scripts download and fly GPS waypoint missions, arm/takeoff in GUIDED mode, log coordinates, and run detection on camera frames. YOLOv5 — trained 4 h on DJI-captured litter photos — scored 90% field accuracy, with a cheap HSV+contour fallback catching litter in stream for real-time frames. Detected coordinates get posted into a Postgres-backed `/api/litter` (lat/lng, status active → picked_up) which an Expo React Native app renders onto a live map. Recovery is delegated to the ROS2 walking robot (see the Walking Robot project).
The outcome
90% detection accuracy in outdoor tests, ~1 m loiter/landing accuracy, 500 g carried with no attitude degradation, and a documented end-to-end system: survey → detect → GPS tag → map → walk and pick up. Published as the Autonomous Litter Detection and Recovery System research paper.