Strawberry Pedicel Harvesting Pipeline Runs on Jetson Orin

What changed

Researchers opened a full pipeline for pedicel targeted strawberry harvesting using imitation learning on a low cost arm. The work is tied to the IROS Agribotics Workshop 2026 program in Pittsburgh. The repository documents 130 human teleoperation demos spanning about 250k frames.

The policy maps wrist camera pixels to joint actions with ACT and a ResNet-50 backbone. Deployment runs on Jetson Orin AGX at roughly 15 to 20 Hz without cloud compute.

Why it matters

Harvesting fragile fruit usually needs custom vision stacks and motion planners. This project shows an end to end visuomotor policy that targets the 1.4 to 2.4 mm pedicel directly. That matters for soft fruit lines where bruising destroys margin.

Edge inference on Orin fits greenhouse and field robots with weak connectivity. Open dataset and checkpoint links on Hugging Face lower the bar for replication studies.

Who is affected

Ag robotics startups, university ag engineering labs, and produce cooperatives testing robotic pickers. ML engineers can compare ACT based imitation against newer VLAs on the same crop. Safety teams should still validate grip force and human handoff zones in each greenhouse layout.

What to do next

Download the repository, inspect the teleop data format, and replay training curves against your own crop lighting conditions. If you deploy, wrap the policy with hardware estops and fruit damage QA sampling before scaling acreage.

What to watch

Watch for peer reviewed publication beyond the workshop abstract. Track whether authors release multi cultivar datasets with different pedicel thickness.

Photo: RDNE Stock project / Pexels

Training loss in the README falls from 2.84 to 0.091 over 100k steps in the reported configuration.

Photo: cottonbro studio / Pexels

Strawberry picking is hard because the stem is thin and the fruit bruises easily. A policy that aims at the pedicel can reduce flesh damage compared with grasping the berry body. Imitation learning here skips explicit 3D models of each berry as long as teleop coverage spans lighting and clutter seen in production rows.

Farm operators should still validate pick rates and damage metrics per cultivar. A model trained on one greenhouse camera height may fail when beds are replanted or when dust covers the lens. Pair ML pickers with human spot checks during the first season.

Open Hugging Face artifacts help academic teams benchmark against this ACT baseline before trying larger VLAs that cost more to run on edge GPUs.

Behavioral cloning means the robot watches human moves and copies them. No hand written rules pick each stem. The camera sees what the human saw. The arm repeats joint motions that matched successful picks in training.

Low cost arms plus edge GPUs fit one row at a time. Cloud free designs struggle when LTE is weak in the field.

Measure bruise rate per hundred picks. A fast policy that damages fruit is not ready. Track both speed and damage in weekly reports.

Share lighting specs with the data team. Greenhouse LEDs differ from open field sun. Models trained in one may fail in the other.

Record row width and bed height in dataset cards. Geometry shifts the camera view. Labels without geometry confuse retrain jobs.

Invite pickers to label failure modes. They see stem breaks humans miss. Those notes improve the next teleop session plan.

Run night and day pick tests if your market ships both. Dew and glare change images. One schedule is not enough.

Budget time to wash cameras each shift. Dust drops success rates fast. Clean lenses beat bigger models here.

Sources

• Primary — Nitin Rai et al., Pedicel harvesting repository (2026). Hosts code, dataset pointers, and deployment metrics for IROS Agribotics Workshop 2026.
• Secondary — Robotics paper daily index, Agricultural VLA paper listing (October 2026). Places agricultural manipulation VLA work in context with other October robotics preprints.