Related Experiment Video
Updated: Aug 30, 2025
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
1.7K
GR-ConvNet v2: A Real-Time Multi-Grasp Detection Network for Robotic Grasping
Sulabh Kumra1,2, Shirin Joshi1,3, Ferat Sahin1
1The Department of Electrical Engineering, Rochester Institute of Technology, Rochester, NY 14623, USA.
Sensors (Basel, Switzerland)
|August 26, 2022
Summary
This study introduces a dual-module robotic system using an improved Generative Residual Convolutional Neural Network (GR-ConvNet v2) for real-time robotic grasp generation. The system achieves state-of-the-art accuracy on multiple datasets and demonstrates high success rates in real-world robotic manipulation tasks.
Area of Science:
- Robotics
- Computer Vision
- Machine Learning
Background:
- Generating robotic grasps for unknown objects is a significant challenge in robotics.
- Current methods often struggle with real-time performance and generalization to novel objects.
Purpose of the Study:
- To develop a robust and efficient system for generating and performing antipodal robotic grasps for unknown objects.
- To improve the accuracy and real-time capabilities of robotic grasp generation using deep learning.
Main Methods:
- A dual-module robotic system incorporating an improved Generative Residual Convolutional Neural Network (GR-ConvNet v2) was proposed.
- The GR-ConvNet v2 model processes n-channel images to generate antipodal grasps at real-time speeds (20 ms).
- The system was evaluated on standard grasping datasets (Cornell, Jacquard, Graspnet) and tested in simulation and real-world scenarios with diverse objects.
Main Results:
- Achieved state-of-the-art accuracy: 98.8% (Cornell), 95.1% (Jacquard), and 97.4% (Graspnet).
- Outperformed prior work using a stricter Intersection over Union (IoU)-based grasp detection metric.
- Demonstrated high grasp success rates on a 7 Degrees of Freedom (DoF) robotic manipulator: 95.4% (household objects) and 93.0% (adversarial objects).
Conclusions:
- The proposed GR-ConvNet v2 model offers a significant advancement in real-time, accurate robotic grasp generation.
- The system exhibits strong adaptability and generalization capabilities, performing well on unseen objects and complex tasks.
- This work paves the way for more capable and versatile robotic manipulation systems in unstructured environments.
