Autonomous Blueberry Harvester

In the field

The platform deployed on a commercial blueberry row — a tracked mobile base carrying the arm, the perception stack, and a collection bag.

The harvesting robot on a tracked base beside a blueberry bush, arm extended with a white collection bag — The harvester working a commercial blueberry row, arm extended with the collection bag deployed.

Demo video

Wide 16:9 clip of a complete harvest cycle. An animated GIF or a poster-framed MP4 both work here.

A full harvest cycle, from cluster detection to detachment.

Overview

A global camera scans the bush and proposes cluster waypoints; the arm presents an in-hand camera for close-range visual servoing; a soft rolling gripper grasps and detaches while watching motor current, so it never crushes the fruit.

Perceive

Vision-guided targeting

YOLO cluster detection, an ArUco-anchored world frame, and PCA orientation estimation feed 3-D harvest waypoints.

Reach

6-DOF kinematics

Analytic-seeded numerical IK solves both position and the radial approach orientation needed for grasping.

Grasp

Force-aware rolling gripper

Current feedback clamps and detaches clusters gently — the core research contribution.

Detector output across a whole blueberry bush, with ripe fruit boxed in magenta and unripe in yellow — Detector output across a full bush — ripe fruit in magenta, unripe in yellow.

Close-range detector output separating ripe and unripe blueberries on a single cluster — At close range, where ripeness decides whether a cluster is picked.

System at a glance

The hardware, as configured in the control pipeline.

Arm	RRP arm (pitch · yaw · prismatic) + wrist (yaw · pitch · roll)
Cameras	Stereolabs ZED global (HD1080) · Intel RealSense in-hand (640×480 + depth)
Gripper	2 grasp fingers + 2 rolling bands (Dynamixel XM) · current sensed at ~3.36 mA/count

Current-based, force-aware grasping

Delicate fruit bruises under low force. Rather than adding a tactile sensor, the gripper infers contact force from Dynamixel motor current — taking that from binary contact detection to calibrated, regulated force control.

Close-up of the rolling gripper end-effector showing the grasp fingers, rolling bands and Dynamixel servos — The end-effector up close: two grasp fingers, two driven rolling bands, and the Dynamixel servos whose present-current reading stands in for a force sensor.

01 · MEASURE

Ground-truth force

Ramp closure against a force sensor; log current vs. force, up and down for hysteresis.

02 · MODEL

Current → force

Fit F = m·I + b; locate the current at a configurable bruise-force threshold.

03 · REGULATE

Hold a gentle F*

A model-based controller drives estimated clamp force to a target in Newtons.

04 · COMPARE

Bruise-risk metrics

Overshoot, settling time and time-above-bruise vs. the threshold controller.

Team

The Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology.

Yixuan Xia
Yilin Cai
Samuel E. Wilcox
Yue Chen

Two members of the project team standing beside the harvesting robot in a blueberry field — Field testing at the research farm.

Citation

If you use this work, please cite it. TODO: confirm the final title, venue, and year.

@misc{blueberry_harvester,
  title  = {Autonomous Blueberry Harvesting with a Force-Aware Rolling Gripper},
  author = {Xia, Yixuan and Cai, Yilin and Wilcox, Samuel E. and Chen, Yue},
  year   = {2026},
  note   = {https://github.com/StopHere1/autonomous_blueberry_harvester}
}

Code, setup instructions and the full experiment toolkit live in the GitHub repository.