AI Weed Detection Robot

A differential-drive field rover that pairs a YOLOv11 weed/crop discriminator with a row-following navigation stack. Frames from a downward-facing camera are classified on a Jetson Nano; detected weed centroids are projected into the spray-boom frame and trigger individually addressable nozzles for centimetre-accurate application. GPS waypoints define the coverage mission while a local costmap handles obstacle avoidance.

Blanket herbicide spraying wastes chemicals, raises costs, and damages soil health — yet manual spot-spraying doesn't scale. The goal is an autonomous rover that distinguishes weeds from crops on-device and sprays only the weeds.

A YOLOv11 model fine-tuned on field imagery runs on a Jetson Nano to separate weeds from crop rows. Detections are projected from image space into the spray-boom frame, and individually addressable nozzles fire per-target. ROS 2 coordinates GPS waypoint following, a LiDAR/vision costmap for obstacle avoidance, and the spray controller.

Planned — to be measured during field trials: weed-detection mAP, spray-targeting accuracy, and percentage reduction in herbicide volume versus blanket spraying.

1. Phase 1

   Dataset & detector

   Collect and label a weed/crop dataset and fine-tune YOLOv11 for on-device inference.

2. Phase 2

   Spray targeting

   Calibrate the camera-to-boom transform and drive individually addressable nozzles from detection centroids.

3. Phase 3

   Navigation

   ROS 2 GPS waypoint following with row detection and costmap-based obstacle avoidance.

4. Phase 4

   Field trials

   Benchmark detection accuracy, spray precision, and chemical savings across crop types and lighting.