Research on Multi-Object Sorting System Based on Deep Learning.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Robot sorting in unstructured environments is a complex challenge.
• Existing systems often lack efficiency, stability, and accuracy for stacked objects.
• Advancements in deep learning offer potential solutions for improved robotic manipulation.

Purpose of the Study:

• To develop a robust robot multi-object sorting system for unstructured scenes.
• To enable robots to perform simple, efficient, stable, and accurate sorting of stacked objects.
• To integrate rotating target detection and instance segmentation for precise object manipulation.

Main Methods:

• Constructed a training model for rotating target detection using object placement data.
• Developed an instance segmentation model, specifically an optimized Mask R-CNN, for object surface segmentation.
• Extracted upper surface point clouds to calculate normal vectors and determine object attitude.
• Fused object posture, category, and grasping sequence for robotic control.
• Tested system performance on an experimental platform, evaluating success rates.

Main Results:

• The trained rotating target detection model accurately identified object position, rotation angle, and category.
• The optimized Mask R-CNN effectively segmented object surfaces and extracted point clouds.
• The integrated system demonstrated efficient, accurate, and stable sorting of stacked objects.
• Experimental results confirmed the system's high success rate in object capture.

Conclusions:

• The proposed deep learning-based multi-object sorting system significantly enhances robotic sorting capabilities in unstructured environments.
• The system provides a stable and accurate solution for handling stacked objects.
• This research contributes to the advancement of intelligent robotic manipulation and automation.