GR-ConvNet

Table of contents

1. Core ideas
2. Typical workflow
3. Manipulation relevance

---

GR-ConvNet (Generative Residual Convolutional Neural Network) predicts antipodal grasp configurations from RGB-D images using a generative residual architecture.

Core ideas

• Image-based grasp generation from depth (and optionally RGB)
• Residual convolutional network for efficient inference
• Outputs grasp quality, angle, and width maps

Typical workflow

1. Capture an RGB-D image of the scene.
2. Run GR-ConvNet to predict grasp quality, orientation, and width.
3. Sample peak-quality pixel locations as grasp candidates.
4. Convert image-plane grasps to robot/gripper poses and execute.

Manipulation relevance

• Fast planar/antipodal grasping from camera views
• Common baseline for learning-based grasp detection
• Complements 6-DoF methods such as GraspNet in cluttered scenes