Welding Robot
lastmod 2026-09-29created 2026-03-095 edits
Categories
Quality assessment using LVS (scanControl)
When handling data from 3D profile sensors such as an LVS (Laser Vision Sensor) in weld surface inspection, the most important question is how to process the raw Point Cloud data before feeding it into an AI model.
Based on the data being collected, here are two AI models and approaches recommended as the most reliable and efficient in practice.
2.5D Depth Map-based inspection (most recommended)
This approach projects the extracted 3D Point Cloud data in PLY form along the Z axis (height), converts it into a 2D grayscale image (Depth Map) where height is represented as pixel brightness (Intensity), and then applies mature 2D Vision AI models.
- Anomaly Detection based on unsupervised learning
- Recommended models: PatchCore, PaDiM
- When to apply: Very useful early on, when defect data is severely lacking and normal weld data is overwhelmingly abundant. After learning only the features of normal weld beads (Bead), it performs excellently at computing defect scores for regions that deviate from them (porosity, undercut, spatter, etc.).
- Segmentation & Detection based on supervised learning
- Recommended models: YOLOv11-Seg, U-Net (or DeepLabV3+)
- When to apply: Use when enough defect data has been labeled and the location and type of defects must be pinpointed at the pixel level. The YOLO family in particular has very fast real-time inference speed, which is an advantage for deployment on production lines.
Direct 3D Point Cloud processing
To avoid the loss of side curvature or fine geometric information that can occur when projecting to a Depth Map, this approach feeds the 3D data itself directly into the model. It is relatively compute-heavy, and optimizing it for real-time inference is tricky.
- Recommended models: PointNet++, MinkowskiEngine (Sparse 3D Convolution)
- When to apply: Suitable when, beyond the presence of defects, the geometric shape itself, such as bead width/height/volume defects, must be finely measured and classified in 3D space.
AI Welding Robot
Notes on "AI-based automatic welding robots".
- AIWeldingRobot:Basic - Basics
See also
- 3D Segmentation
- Welding
- ROS2
- RTOS
- Machine Vision Camera
- AI/LLM/Deep Learning/Machine Learning
- Point Cloud
- RoboDK