A Dual-Modal Robot Welding Trajectory Generation Scheme for Motion Based on Stereo Vision and Deep Learning

Area of Science:

• Robotics
• Computer Vision
• Deep Learning

Background:

• Existing teaching-free methods struggle with redundant point cloud processing and robustness in complex environments.
• Challenges include accurate weld segmentation and feature point extraction for robot path planning.

Purpose of the Study:

• To propose a novel dual-modal perception framework integrating 2D image recognition and 3D point cloud planning.
• To enhance the autonomy and robustness of robot trajectory planning for welding applications.

Main Methods:

• Developed an improved U-Net deep learning model with VGG16 backbone and dual-channel attention module (DAM) for weld segmentation.
• Utilized Zhang-Suen algorithm for weld centerline extraction and polynomial fitting for feature point identification.
• Implemented an improved RANSAC algorithm and equal-area weld bead filling strategy for groove feature extraction and trajectory planning.

Main Results:

• Achieved high weld segmentation performance with mIoU of 0.887 and F1-Score of 0.940.
• Obtained precise feature point positioning with a mean absolute error (MAE) of 0.238 mm.
• Demonstrated system accuracy of 0.208 mm and weld width fluctuation within ±0.15 mm under complex conditions.

Conclusions:

• The proposed dual-modal framework significantly improves the autonomy and robustness of robot trajectory planning.
• The method maintains high accuracy even with noise interference and groove deformation, crucial for industrial applications.