Depth Image-Based Deep Learning of Grasp Planning for Textureless Planar-Faced Objects in Vision-Guided Robotic...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Bin-picking textureless objects is challenging for traditional vision systems due to difficulties in feature extraction and the impracticality of preparing numerous goal images.
• Existing methods often struggle with objects lacking distinct visual features, limiting automation in warehouse logistics.

Purpose of the Study:

• To develop a robust depth image-based vision-guided robot bin-picking system for textureless, planar-faced objects.
• To overcome limitations in feature extraction for textureless objects and the impracticality of manual goal image preparation.

Main Methods:

• Utilized a deep convolutional neural network (DCNN) trained on 15,000 synthetic depth images to directly predict grasp points without object segmentation.
• Developed a DCNN capable of predicting optimal grasp patterns for a two-vacuum-cup suction hand.
• Introduced a surface feature descriptor to refine grasp point positions using surface features (center position and normal), eliminating the need for texture features.

Main Results:

• The proposed system demonstrated high efficiency in picking randomly posed textureless boxes in cluttered environments.
• Achieved a 97.5% success rate with a 7-degrees-of-freedom robot.
• Reached picking speeds exceeding 1000 pieces per hour.

Conclusions:

• The novel depth image-based approach effectively addresses the challenges of bin-picking textureless objects.
• The system eliminates the need for texture features and extensive sim-to-real modification, offering a practical solution for warehouse automation.
• The developed method significantly enhances robotic picking capabilities for common industrial tasks.