Computer Vision · Kantor Lab, CMU

Catching crop disease in the field

Real-time YOLO detection on a farm-ng rover

An autonomous rover drives the orchard and spots fireblight as it goes, flagging infected trees for precision pruning. Won the farm-ng Challenge and a $25K grant.

0.87 detection mAP
30 FPS on-robot
$25K grant, farm-ng winner

Detection that runs while the robot drives

Research at CMU's Kantor Lab. Manual scouting can't scan orchards fast enough to contain fireblight outbreaks. I built a real-time YOLO detection pipeline on ROS2 aboard an autonomous farm-ng rover that classifies infection severity on-robot and drops GPS-tagged markers, running live under bad lighting, motion blur, and occlusion. It won the farm-ng Challenge 2026 and a $25K grant.

Fig 1 — The model locks onto infected foliage in real time as the rover moves through the rows.
01

Real-time on ROS2

YOLO detection running live on the farm-ng rover during field traversal, not post-hoc on saved footage.

02

Robust to the field

Active-learning data curation and photometric/geometric augmentation to survive lighting, blur, and occlusion.

03

Edge inference

TensorRT-quantized (INT8) models for low-latency inference on the robot's onboard compute.

04

Actionable output

Severity classification with GPS-tagged markers so growers can prune the right trees.

Fig 2 — Small-lesion detection mAP, before vs after data curation and augmentation.
0.87
mAP, up from 0.61
30 FPS
On-robot inference
$25K
farm-ng Challenge grant
1st
farm-ng Challenge 2026
YOLO ROS2 OpenCV TensorRT Edge Inference Active Learning Data Augmentation Python Robot Autonomy