Related Experiment Video
Updated: Jan 9, 2026
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
2.1K
GR-AttNet: Robotic grasping with lightweight spatial attention mechanism
Shengbang Zhou1, Zhongsheng Liu1, Chuanqi Li1
1Guangxi Key Laboratory of Functional Information Materials and Intelligent Information Processing, Nanning Normal University, Nanning, China.
Plos One
|December 4, 2025
Summary
This study introduces the Generative Residual Attention Network (GR-AttNet) for robotic grasping, improving accuracy and efficiency in cluttered environments. GR-AttNet demonstrates superior performance on benchmark datasets, offering a practical solution for complex robotic tasks.
Area of Science:
- Robotics
- Computer Vision
- Artificial Intelligence
Background:
- Robotic grasping is essential for automation in manufacturing, logistics, and service industries.
- Current grasping methods face challenges with object occlusion and complex arrangements in cluttered scenes.
- There is a need for more robust and efficient robotic grasping solutions.
Purpose of the Study:
- To propose a novel deep learning model, the Generative Residual Attention Network (GR-AttNet), for enhanced robotic grasping.
- To improve the adaptability and efficiency of robotic grasping systems in cluttered environments.
- To achieve high accuracy and speed in grasp detection.
Main Methods:
- The Generative Residual Attention Network (GR-AttNet) was developed, building upon the Generative Residual Convolutional Neural Network (GR-CNN).
- A spatial attention mechanism was incorporated to enhance adaptability to cluttered scenes.
- Architectural optimization was performed to reduce parameters and improve efficiency.
Main Results:
- GR-AttNet achieved 98.1% accuracy on the Cornell Grasping Dataset and 94.9% on the Jacquard Dataset.
- Processing speeds were recorded at 148 ms and 163 ms, respectively, outperforming state-of-the-art methods.
- The model, with only 2.8 million parameters, demonstrated over 50% success rates in complex simulated scenarios.
Conclusions:
- GR-AttNet presents a novel and highly practical solution for robotic grasping tasks.
- The model shows significant improvements in accuracy, efficiency, and adaptability compared to existing methods.
- Future work may focus on addressing challenges in small object detection and severe occlusion cases.
