Scholarlink

<p>Accurate and real-time object grasping in dynamic environments is a key challenge in industrial automation, particularly for small- and medium-scale applications where cost and computational efficiency are critical. The existing robotic systems often depend on expensive industrial manipulators and GPU-based processing for object detection and 3D localization, limiting their scalability and practical deployment. This paper presents a 3D vision-based robotic system with low-cost industrial implementation for automated object grasping, combining the YOLOv12 deep learning model with a 3D Intel RealSense D435 camera. A custom dataset comprising 10 objects, such as cartons, bottles, cans, and pouches, was collected and annotated under diverse environmental conditions to improve detection robustness. The YOLOv12 model, trained on this custom dataset, achieved a high detection precision of 98.6%, with a mean average precision (mAP50) of 91.5% and mAP50-95 of 72.3%, demonstrating strong generalization across various object types and orientations. Depth data captured by the D435 camera were enhanced using a sequence of spatial, temporal, and disparity filter techniques. As a result, the system achieved 3D localization errors of about 3.22&#xa0;mm along the Z-axis, and standard deviations below 2.13&#xa0;mm across all spatial dimensions. A four-degree-of-freedom SCARA robotic arm, actuated by stepper motors and controlled through a Delta PLC over RS485 communication, was deployed to execute the grasping operation. The integrated system achieved a real-time inference speed of about 16&#xa0;ms per frame and a practical pick-and-place productivity of 20 items per minute. Experimental results demonstrate the system’s accuracy, speed, and adaptability for real-world industrial automation applications.</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

Real-time 3D vision-based robotic grasping system for low-cost industrial production lines

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Vo Duy Cong,
  • Le Hoai Phuong,
  • Phan Xuan Trung

摘要

Accurate and real-time object grasping in dynamic environments is a key challenge in industrial automation, particularly for small- and medium-scale applications where cost and computational efficiency are critical. The existing robotic systems often depend on expensive industrial manipulators and GPU-based processing for object detection and 3D localization, limiting their scalability and practical deployment. This paper presents a 3D vision-based robotic system with low-cost industrial implementation for automated object grasping, combining the YOLOv12 deep learning model with a 3D Intel RealSense D435 camera. A custom dataset comprising 10 objects, such as cartons, bottles, cans, and pouches, was collected and annotated under diverse environmental conditions to improve detection robustness. The YOLOv12 model, trained on this custom dataset, achieved a high detection precision of 98.6%, with a mean average precision (mAP50) of 91.5% and mAP50-95 of 72.3%, demonstrating strong generalization across various object types and orientations. Depth data captured by the D435 camera were enhanced using a sequence of spatial, temporal, and disparity filter techniques. As a result, the system achieved 3D localization errors of about 3.22 mm along the Z-axis, and standard deviations below 2.13 mm across all spatial dimensions. A four-degree-of-freedom SCARA robotic arm, actuated by stepper motors and controlled through a Delta PLC over RS485 communication, was deployed to execute the grasping operation. The integrated system achieved a real-time inference speed of about 16 ms per frame and a practical pick-and-place productivity of 20 items per minute. Experimental results demonstrate the system’s accuracy, speed, and adaptability for real-world industrial automation applications.