A Deep Learning-Based Framework for Offline Robotic Weld Path Generation Using a Single Top-View RGB-D Image

Area of Science:

• Robotics
• Computer Vision
• Materials Science

Background:

• Automated pipe welding demands precise weld seam extraction and robotic path generation, especially for complex geometries.
• Current vision-based methods often involve costly laser systems, multi-view setups, or continuous tracking, increasing complexity and expense.

Purpose of the Study:

• To develop a deep learning-based offline robotic welding framework for generating 3D welding paths from a single top-view RGB-D image.
• To address limitations of existing methods by reducing hardware costs and system complexity.

Main Methods:

• A unified pipeline integrating weld seam detection, semantic segmentation, morphology-based post-processing, RGB-D image alignment, coordinate transformation, and polynomial trajectory refinement.
• Development of a custom pipe welding dataset with 1476 annotated images of industrial pipe materials.
• Evaluation using metrics such as mean Intersection over Union (mIoU), Recall, mean Average Precision (mAP50), and Root Mean Square Error (RMSE).

Main Results:

• The Region of Interest (ROI)-guided pipeline improved U-Net segmentation mIoU from 0.735 to 0.791.
• The detection model achieved a Recall of 0.988 and mAP50 of 0.995.
• Polynomial trajectory refinement reduced 3D positional RMSE to 0.333 mm, enabling continuous welding without path modification.

Conclusions:

• The proposed framework offers a practical and cost-effective solution for offline robotic weld seam extraction and path generation.
• The study lays a foundation for future extensions towards online robotic welding with real-time tracking and adaptive correction.