Scholarlink

<p>To address the challenges of automated detection and grasping caused by the extremely small scale, irregular morphology, and severe metal reflection of bearing defects, this study proposes a vision-guided closed-loop system using the YOLO_grab algorithm and a UR5 robotic arm. To overcome the perception bottleneck for tiny targets(linear depression, groove, abrasion and scratch), the baseline YOLOv8 is reconstructed: the Expectation-Maximization Attention (EMA) mechanism suppresses background noise and enhances focus on low-contrast features; the Asymptotic Fractional Pyramid Network (AFPN) eliminates semantic gaps in cross-scale feature fusion; and the SIoU loss function decouples angle and distance penalties, resolving the localization oscillation and size distortion of bounding boxes. Experiments demonstrate YOLO_grab achieves a 99.50% mAP@0.5 and an 88.70% mAP@0.5:0.95, surpassing SOTA models like YOLOv26. Furthermore, a rigorous Eye-to-Hand coordinate transformation model effectively blocks visual error propagation to the physical execution layer. Dynamic grasping tests show an inference speed of 142 FPS with a 7 ms single-frame latency, far exceeding real-time control requirements. In Sim-to-Real deployment, the system stably maintains a sorting efficiency of 15 grasps per minute (GPM). With minimal speed compromise, this system integrates high-fidelity visual perception and precise 3D physical control, providing a highly robust solution for industrial defect sorting.</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

A YOLO_grab-based grasping system for defective bearings using a UR5 robotic arm

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Zhaoxuan Li,
  • Mohd Salman Abu Mansor

摘要

To address the challenges of automated detection and grasping caused by the extremely small scale, irregular morphology, and severe metal reflection of bearing defects, this study proposes a vision-guided closed-loop system using the YOLO_grab algorithm and a UR5 robotic arm. To overcome the perception bottleneck for tiny targets(linear depression, groove, abrasion and scratch), the baseline YOLOv8 is reconstructed: the Expectation-Maximization Attention (EMA) mechanism suppresses background noise and enhances focus on low-contrast features; the Asymptotic Fractional Pyramid Network (AFPN) eliminates semantic gaps in cross-scale feature fusion; and the SIoU loss function decouples angle and distance penalties, resolving the localization oscillation and size distortion of bounding boxes. Experiments demonstrate YOLO_grab achieves a 99.50% mAP@0.5 and an 88.70% mAP@0.5:0.95, surpassing SOTA models like YOLOv26. Furthermore, a rigorous Eye-to-Hand coordinate transformation model effectively blocks visual error propagation to the physical execution layer. Dynamic grasping tests show an inference speed of 142 FPS with a 7 ms single-frame latency, far exceeding real-time control requirements. In Sim-to-Real deployment, the system stably maintains a sorting efficiency of 15 grasps per minute (GPM). With minimal speed compromise, this system integrates high-fidelity visual perception and precise 3D physical control, providing a highly robust solution for industrial defect sorting.