Cascaded Feature Fusion Grasping Network for Real-Time Robotic Systems

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic grasping of irregular objects is a significant challenge.
• Accurate and efficient grasping pose estimation is crucial for robotic applications.

Purpose of the Study:

• To propose a novel RGB-D data-based grasping pose prediction network, the Cascaded Feature Fusion Grasping Network (CFFGN).
• To achieve high-efficiency, lightweight, and rapid grasping pose estimation.

Main Methods:

• Utilized depth-wise separable convolutions for efficiency.
• Incorporated convolutional block attention modules for feature focus.
• Employed multi-scale dilated convolution for expanded receptive fields.
• Implemented bidirectional feature pyramid modules for multi-level feature fusion.

Main Results:

• Achieved 66.7 frames per second grasping pose prediction speed on the Cornell dataset.
• Obtained 98.6% accuracy (image-wise) and 96.9% accuracy (object-wise) on the Cornell dataset.
• Demonstrated an average grasping success rate of 95.6% in real-world experiments with parallel grippers.

Conclusions:

• The CFFGN network effectively balances high-speed processing with high accuracy in grasping pose prediction.
• The proposed method shows significant promise for real-world robotic grasping applications.