Learning Suction Graspability Considering Grasp Quality and Robot Reachability for Bin-Picking.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Deep learning is crucial for robotic grasping, but human-labeled datasets are costly.
• Existing methods often use complex physical models for grasp evaluation, requiring extensive parameter tuning.
• Previous approaches overlooked manipulator reachability, limiting grasp success in real-world scenarios.

Purpose of the Study:

• To develop an intuitive, geometric analytic-based grasp quality evaluation metric.
• To integrate a reachability evaluation metric for enhanced grasp planning.
• To train a deep learning model (SG-U-Net++) for pixel-wise grasp quality and reachability assessment.

Main Methods:

• Generated synthetic images using a physical simulator.
• Annotated images with a novel geometric analytic grasp quality metric and a reachability metric.
• Trained a suction graspability U-Net++ (SG-U-Net++) auto-encoder-decoder model on annotated synthetic data.

Main Results:

• The proposed geometric analytic metric demonstrates competitive performance against physically-inspired metrics.
• Incorporating reachability evaluation significantly reduces motion planning computation time.
• The SG-U-Net++ system achieved a high picking speed of 560 pieces per hour.

Conclusions:

• The novel geometric analytic grasp quality metric is effective and practical.
• Integrating reachability assessment enhances the efficiency of robotic grasping systems.
• The SG-U-Net++ model offers a robust solution for automated robotic picking.