Scholarlink

<p>Robotic welding requires the ability to track paths accurately and smoothly in dynamic environments. However, traditional methods suffer from sensor calibration issues and the creation of specific paths. We propose a hybrid framework that integrates YOLOv8-seg for real-time welding path segmentation with a multi-stage hierarchical Soft Actor Critic algorithm. The YOLOv8-seg model was trained on a custom dataset of 450 welding path images without data augmentation, achieving a mAP of 95%. The full 3D path coordinates (X, Y, Z) are extracted via RGB-D masks and Zhang-Suen skeletonization, completely eliminating the need for laser sensors and reducing costs. The MS-SAC agent divides the path-tracking tasks hierarchically so that the first stage classifies open/closed paths and selects the welding starting point, the second stage specializes in tracking a specific type of path, and the third stage generalizes the paths and movement patterns, resulting in faster convergence than standard SAC. Simulation results within CoppeliaSim using a KUKA KR8 6-DOF robot show an average max error of 0.5&#xa0;mm across different paths. This method enhances the scalability and generalization of Industry 4.0 and paves the way for more advanced robotic welding systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid vision-hierarchical soft actor-critic approach for autonomous robotic welding

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Mohamed Yasser Korichi,
  • Amira Chekir

摘要

Robotic welding requires the ability to track paths accurately and smoothly in dynamic environments. However, traditional methods suffer from sensor calibration issues and the creation of specific paths. We propose a hybrid framework that integrates YOLOv8-seg for real-time welding path segmentation with a multi-stage hierarchical Soft Actor Critic algorithm. The YOLOv8-seg model was trained on a custom dataset of 450 welding path images without data augmentation, achieving a mAP of 95%. The full 3D path coordinates (X, Y, Z) are extracted via RGB-D masks and Zhang-Suen skeletonization, completely eliminating the need for laser sensors and reducing costs. The MS-SAC agent divides the path-tracking tasks hierarchically so that the first stage classifies open/closed paths and selects the welding starting point, the second stage specializes in tracking a specific type of path, and the third stage generalizes the paths and movement patterns, resulting in faster convergence than standard SAC. Simulation results within CoppeliaSim using a KUKA KR8 6-DOF robot show an average max error of 0.5 mm across different paths. This method enhances the scalability and generalization of Industry 4.0 and paves the way for more advanced robotic welding systems.