💻 computer science

ROG-Grasp: Root-Oriented Geometry for Robotic Grasping and Placement

This paper introduces ROG-Grasp, a geometry-based robotic framework that leverages RGB-D perception and root surface analysis to achieve reliable, orientation-aware grasping and placement of agricultural produce, demonstrating superior accuracy and speed compared to vision-language-action policies.

Original authors: Zijian An, Augustus Sroka, Ran Yang, Bill Cai, Satoru Eto, Brian Poon, Kelvin Cai, Shijie Geng, Feng Liu, Yiming Feng, Lifeng Zhou

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zijian An, Augustus Sroka, Ran Yang, Bill Cai, Satoru Eto, Brian Poon, Kelvin Cai, Shijie Geng, Feng Liu, Yiming Feng, Lifeng Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a robot chef trying to pack a box of fresh vegetables. You can easily pick up a tomato or an onion, but here's the tricky part: you need to place them in the box standing up, with their roots pointing down, just like they would sit in a garden. If you drop them randomly, they might roll over, get bruised, or look messy in the store display.

This paper introduces a new "brain" for robots called ROG-Grasp that solves this specific problem. Instead of guessing or trying to learn through trial and error, this system uses simple geometry to figure out exactly how to hold and place the produce.

Here is how it works, broken down into everyday concepts:

1. The "Root Detective" (Seeing the Vegetable)

Most robots just look at the whole vegetable. But ROG-Grasp has a special trick: it looks specifically for the root or the stem area.

The Analogy: Think of it like a detective looking for a specific clue. The robot uses a camera (like a human eye) and a special depth sensor (like a 3D scanner) to find the root.
The Tool: It uses a smart software tool called YOLO (which is like a super-fast photo sorter) to spot the root area on the tomato or onion.

2. The "Flat Surface" Trick (Figuring Out the Angle)

Once the robot finds the root, it needs to know which way is "up."

The Analogy: Imagine the root of an onion or a tomato is like a tiny, flat table top. Even though the vegetable is round, that little root patch is relatively flat.
The Math: The robot takes a bunch of 3D points from that root area and draws an invisible flat sheet (a plane) through them. The direction this sheet is facing tells the robot exactly how the vegetable is tilted. It's like using a level tool to see if a picture frame is crooked.

3. The "Handshake" (Grabbing and Placing)

Now that the robot knows the angle, it calculates the perfect way to grab the vegetable.

The Approach: It doesn't just dive straight down. It follows a pre-planned dance of six steps (waypoints). It approaches from above, grabs the vegetable, lifts it up, and then carefully rotates it so the root points down.
The Placement: It has a "target box" (the holder) where it needs to put the vegetable. The robot moves its hand to match the angle of the box perfectly, ensuring the vegetable lands standing up, not on its side.

4. The Race: Geometry vs. "Learning"

The authors compared their new method (ROG-Grasp) against a popular type of AI called VLA (Vision-Language-Action).

The VLA Robot: This robot tries to "learn" how to do the task by watching videos and guessing, kind of like a student trying to solve a math problem by guessing numbers until they get it right.
The ROG-Grasp Robot: This robot uses the "flat root" math trick. It knows the answer immediately because it measures the geometry.

The Results:

Speed: The ROG-Grasp robot was much faster (about 8 seconds vs. 20+ seconds for the learning robot).
Success: The ROG-Grasp robot succeeded 80–90% of the time, even when there were other vegetables in the way. The learning robot only succeeded about 10–40% of the time, often getting confused or dropping the vegetables.
Why? The learning robot struggled to find the root and get the angle right, while the geometry-based method was precise and reliable.

The Bottom Line

This paper shows that for specific jobs like packing vegetables neatly, you don't always need a robot that tries to "think" like a human. Sometimes, a robot that uses simple, reliable math to measure the shape of a root is much faster, more accurate, and better at getting the job done without making a mess.

Technical Summary: ROG-Grasp

Problem Statement

While robotic systems have advanced significantly in agricultural harvesting and detection, a critical gap remains in post-harvest processing and packaging. In modern packhouses, produce must often be sorted and placed into containers with consistent, specific orientations to facilitate downstream processing, inspection, and commercial presentation. Existing robotic solutions typically focus on grasping and depositing objects without explicitly controlling their final pose. Furthermore, while Vision-Language-Action (VLA) models have shown promise in general manipulation, preliminary investigations indicate they struggle to reliably identify specific root regions of agricultural produce, leading to inaccurate orientation estimation and unstable grasping in cluttered environments.

Methodology

The paper proposes ROG-Grasp, a geometry-based framework designed to estimate produce orientation from root surface geometry using RGB-D perception. The system operates through two primary modules: Visual Perception and Motion Execution.

1. Visual Perception

Root Detection: A YOLO-based segmentation model is fine-tuned to detect both the produce body and the specific root region. The training dataset is constructed using a semi-automatic pipeline where body masks are generated via foreground-background contrast, while root annotations are manually created.
3D Point Cloud Extraction: Detected root regions are projected into 3D space using depth data and camera intrinsics to generate a point cloud in the robot's base frame.
Orientation Estimation: Leveraging the observation that root surfaces (e.g., on tomatoes, onions, garlic) exhibit approximately planar geometry with low curvature, the system applies Least-Squares Plane Fitting to the root point cloud.
- Singular Value Decomposition (SVD) is performed on the centered point cloud.
- The eigenvector corresponding to the smallest singular value defines the root normal vector ( $\mathbf{n}$ ).
- This normal vector serves as the primary cue for the object's orientation.

2. Motion Planning and Control

Grasp Pose Generation: The grasp position is defined as the centroid of the root point cloud ( $\mathbf{c}$ ) plus a predefined offset ( $\delta$ ) along the normal vector ( $\mathbf{n}$ ). The grasp orientation aligns the gripper's Z-axis with $\mathbf{n}$ .
Placement Pose: A fixed "holder frame" defines the desired upright placement configuration. The robot aligns the gripper frame with this holder frame to release the object.
Execution Pipeline: To ensure safety and avoid collisions, the robot executes a sequence of six intermediate waypoints ( $wp_0$ $w p_{0}$ to $wp_5$ $w p_{5}$ ):
1. Observation: Capture scene and detect root.
2. Approach: Move to a waypoint above the target, then descend vertically.
3. Retreat: Move upward after grasping to clear surrounding objects.
4. Reorientation: Adjust orientation to ensure the root faces downward for placement.
5. Placement: Move to the holder and release.
6. Return: Retreat to the initial pose.
Coordinate Transformations: The system utilizes homogeneous transformation matrices in $SE(3)$ to manage transformations between the camera, gripper, connector, and base frames, ensuring precise Cartesian motion commands are generated via Inverse Kinematics (IK).

Key Contributions

Geometry-Driven Orientation Estimation: A novel method that utilizes root surface geometry (via plane fitting on point clouds) to estimate the orientation of agricultural products, enabling orientation-aware manipulation without relying on complex learning-based pose estimators.
Integrated Perception-Action Pipeline: A vision-guided strategy combining YOLO-based root segmentation with 3D geometric fitting to compute stable grasp poses and orientation-constrained Cartesian motion plans.
Empirical Validation: Comprehensive experiments on tomatoes and onions in both isolated and cluttered scenarios, demonstrating the method's robustness compared to state-of-the-art VLA policies.

Experimental Results

The system was evaluated using a Fairino FR5 manipulator with an RGB-D camera on tomatoes and onions.

Success Rates: ROG-Grasp achieved high success rates: 85% (single) and 80% (multi) for tomatoes, and 90% (single) and 80% (multi) for onions.
Comparison with VLA: In contrast, a VLA policy (based on the $\pi_0$ model) achieved significantly lower success rates (40% for single tomato, 10% for multi-tomato) and struggled in cluttered scenes, often failing to infer correct placement orientations.
Execution Time: ROG-Grasp demonstrated significantly faster execution times (~7–8 seconds) compared to the VLA policy (~20–26 seconds), attributed to the VLA's closed-loop action chunking and the geometry-based method's efficient open-loop trajectory planning.
Failure Modes: The primary failure mode for ROG-Grasp occurred during the open-loop descent phase where unintended contact could disturb the object. The VLA policy, while closed-loop, suffered from oscillation and instability in clutter.

Significance and Claims

The paper claims that geometry-driven perception offers a more reliable and efficient solution for orientation-controlled manipulation in agricultural settings compared to end-to-end learning approaches like VLA. By explicitly modeling the planar nature of root surfaces, ROG-Grasp achieves stable grasping and precise placement without the training data burdens and inference latency associated with deep learning policies. The authors emphasize that this approach provides a practical, robust solution for tasks requiring specific object layouts, such as high-value produce packaging, where visual organization and consistent orientation are critical for commercial presentation and protection.

Future work is identified as incorporating closed-loop visual feedback during the grasp phase to handle unexpected object motion and integrating geometry-based estimation with learning-based policies for highly unstructured environments.