GreenSeg Ground Segmentation Tested for Greenhouse Robots in October

What changed

Researchers published GreenSeg, a ground segmentation method for agricultural robots in Mediterranean greenhouses. The ros2 greenseg repository documents tests completed on 5 October 2026 on ROS 2 Humble and Ubuntu 22.04 LTS. The work uses RGB D point clouds to find safe floor paths between crop rows.

GreenSeg combines global plane fitting with a curvature filter and region growing from seed points. That dual layer check aims to keep the navigable plane stable when lighting shifts through the day. The team validated the method on the AGRICOBIOT I platform across four daytime scenarios.

Reported gains include up to 11.58 percent mean recall and 19.24 percent mIoU improvements over benchmark segmenters during end of corridor turns. Those moves matter when robots rotate in tight greenhouse aisles.

Why it matters

Greenhouse operators face tight budgets and changing sun angles. A segmentation stack that holds up across diurnal light reduces manual teleop and stuck events. That directly affects labor cost per hectare.

Open ROS 2 packages lower the bar for integrators who already run humble on field computers. Teams can swap lidar or depth cameras yet keep the same ground plane logic if calibration holds.

Competing methods that assume flat outdoor fields often fail on elevated gutters and cloth shade. GreenSeg targets that structured clutter explicitly.

Who is affected

Greenhouse automation startups. University ag robotics groups in Spain and the EU. Farm managers piloting autonomous carts for harvest assist. Insurers who price liability for slow moving robots near workers.

Sensor vendors selling RGB D kits for horticulture should note the October 2026 test date when quoting support windows.

What to do next

Clone the ros2 greenseg repo and replay bag files on your depth rig. Compare recall on your worst sun glare hour before you commit to a fleet pilot.

Tune seed points for your row spacing. Region growing assumes spatial continuity that may differ in vine crops versus leafy greens.

What to watch

Watch for peer review status on the linked arXiv preprint. Watch maintainer updates to ROS 2 iron or newer distros if your fleet standard moves off humble.

Photo: Quang Nguyen Vinh / Pexels

Preprint results may not match your soil moisture and misting schedule. Budget a season of shadow logging before you remove human spotters.

Photo: Pixabay / Pexels

Keep human estops and geofences even when segmentation scores look strong on paper tests. Greenhouses still host workers and loose tools in lane edges.

Log depth dropout rates when leaves wet the floor. Water films can break plane fits until drains clear.

Pair GreenSeg with low speed caps in shared aisles. Good maps still need time to stop on slippery poly floors.

Invite pickers to mark trouble tiles in a shared map layer. Human tags speed up the next seed point tune pass.

Budget spare depth sensors for dust and spray fouling. Greenhouse air loads differ from open field dust models.

Run weekly safety drills where the robot must stop when a person steps into the lane. Segmentation is only one layer of safety.

Record sun angle and shade cloth state in each test log. GreenSeg claims hinge on diurnal light swings you can replay later.

Share failed runs with the authors through issues, not only wins. Open repos improve when labs post hard frames with consent.

Compare power draw on your PC when running ROS nodes full time. Field boxes need cooling plans in hot summers.

Plan winter tests too. Heating pipes change floor heat patterns that plane fitters must see at least once.

Label each bag file with crop variety and row width. Metadata makes cross farm compares fair later.

Sources

• Primary — University of Almería ARM group, ros2 greenseg repository (5 October 2026). Algorithm description, test date, and ROS 2 integration notes.
• Secondary — arXiv, GreenSeg preprint (2026). Experimental setup and reported segmentation metrics.