A pushing-grasping collaborative method based on deep Q-network algorithm in dual viewpoints.

Area of Science:

• Robotics and Intelligent Systems
• Computer Vision
• Artificial Intelligence

Background:

• Traditional 2D camera-based robot grasping methods suffer from low efficiency and accuracy due to missing information in stacked and occluded scenes.
• Single-view methods cannot actively alter challenging grasping environments, limiting their applicability.

Purpose of the Study:

• To develop a pushing-grasping collaborative method using dual viewpoints to overcome the limitations of single-view robot grasping.
• To enhance the robot's ability to handle complex, occluded, and stacked object scenarios in intelligent manufacturing.

Main Methods:

• An improved deep Q-network algorithm was employed, utilizing an RGB-D camera for dual-viewpoint RGB images and point cloud data.
• The method integrates pushing and grasping actions within the deep Q-network framework, enabling active scene exploration.
• A piecewise reward function was introduced to accelerate the convergence of the deep Q-network training process.

Main Results:

• The proposed dual-view pushing-grasping method achieved a grasping success rate of 83.5% in unstructured scenes within the V-REP simulation environment.
• The approach demonstrated rapid convergence during training, outperforming traditional methods.
• The system exhibited strong generalization capabilities, performing well with novel, previously ungrasped objects.

Conclusions:

• The dual-view pushing-grasping collaborative method significantly improves robot grasping performance in complex and occluded environments.
• The integration of active exploration and improved deep Q-learning offers a robust solution for intelligent manufacturing grasping tasks.
• This approach enhances both the efficiency and accuracy of robotic manipulation in real-world applications.