Related Experiment Video
Updated: Jun 21, 2025
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
1.6K
A Real-Time Grasping Detection Network Architecture for Various Grasping Scenarios
Summary
This study introduces a new neural network for robot grasping detection, enabling robots to grasp unknown objects in real-time. The system achieves high accuracy in simulations and real-world experiments.
Area of Science:
- Robotics
- Computer Vision
- Artificial Intelligence
Background:
- Robot grasping is challenging due to object variability (shape, color, material, pose).
- Existing robot grasping technologies have limitations in handling diverse, unknown objects.
Purpose of the Study:
- To develop an integrated robotic system for grasping numerous unknown objects from alpha-channel images.
- To propose a lightweight, object-independent neural network for real-time grasping detection.
Main Methods:
- Introduced a generative adaptive residual depthwise separable convolutional neural network (GARDSCN).
- GARDSCN processes alpha-channel images for pixel-level grasping detection.
- Network achieves an inference speed of approximately 28 ms for real-time application.
Main Results:
- Achieved 98.88% grasp detection accuracy on the Cornell dataset.
- Achieved 95.23% grasp detection accuracy on the Jacquard dataset.
- Demonstrated a 96.67% grasp success rate in single-object scenes and 94.10% in cluttered scenes on a Kinova Gen2 robot.
Conclusions:
- The proposed GARDSCN effectively addresses grasping detection for unknown objects with varying properties.
- The system demonstrates high accuracy and success rates in both simulated and real-world robotic grasping tasks.
- This research advances real-time grasping detection capabilities for robots in complex environments.
