A two-stage grasp detection method for sequential robotic grasping in stacking scenarios.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Dexterous grasping is crucial for robotic fine manipulation but challenging in stacking scenarios.
• Existing methods struggle with sequential grasping of stacked objects.

Purpose of the Study:

• To propose a novel two-phase approach for grasp detection in sequential robotic grasping for stacking tasks.
• To enhance the accuracy and success rate of robotic grasping in complex stacked environments.

Main Methods:

• Developed a rotated-YOLOv3 (R-YOLOv3) model for detecting top-layer objects in stacked scenarios.
• Created a stacked scenario dataset for training and testing the R-YOLOv3 network.
• Utilized a G-ResNet50 model to determine optimal grasping poses for the uppermost objects.

Main Results:

• The R-YOLOv3 model achieved an average grasping prediction success rate of 96.60% on the Cornell grasping dataset.
• In real-world experiments, the robot achieved a maximum grasping success rate of 95.00% and an average handling success rate of 83.93% in stacked scenarios.

Conclusions:

• The proposed two-phase approach effectively enables robots to perform sequential grasping in complex stacked environments.
• The methodology demonstrates high efficacy and competitiveness for robotic manipulation tasks involving stacked objects.