数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (1): 195-206.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.01.016

doi: 10.11871/jfdc.issn.2096-742X.2026.01.016

• 技术与应用 • 上一篇    下一篇

MSGS: 基于深度相机的多规格机器人抓取系统

付兰慧1(),谯未来1,邓辅秦1,周磊2,*(),侯健峰1,冷雄伟3,周游2,黄广俊2   

  1. 1.五邑大学，电子与信息工程学院， 广东 江门 529020
    2.人工智能与数字经济广东省实验室（深圳），广东 深圳 518132
    3.东莞市李群自动化技术有限公司，广东 东莞 523000
  • 收稿日期:2025-03-02 出版日期:2026-02-20 发布日期:2026-02-02
  • 通讯作者: 周磊
  • 作者简介:付兰慧，博士，五邑大学，讲师，主要研究方向为机器学习，基于机器视觉的检测。
    本文中负责论文撰写与实验。
    FU Lanhui, PhD, is a lecturer at Wuyi University. Her main research interests include machine learning and machine vision-based detection.
    In this paper, she is mainly responsible for paper writing and experimental demonstration.
    E-mail: j002886@wyu.edu.cn|周磊，博士，人工智能与数字经济广东省实验室（深圳）副研究员，高级工程师，研究方向为机器视觉与控制工程。
    本文中主要负责文章方向提出和内容修改。
    ZHOU Lei, Ph.D., is an associate researcher at Guangdong Provincial Laboratory of Artificial Intelligence and Digital Economy (Shenzhen). He is also a senior engineer, and his research interests include machine vision and control engineering.
    In this paper, he is mainly responsible for proposing the direction and revising the manuscript.
    E-mail: zhoulei@gml.ac.cn
  • 基金资助:
    五邑大学博士科研启动经费(BSQD2222);国家重点研发计划“战略性科技创新合作”重点专项-半导体器件封装质量智能检测关键技术研究与应用示范(2023YFE0205800);五邑大学港澳联合基金项目-多模态表征学习下的自监督视频行为分析理论研究与机器智能应用(2022WGALH17)

MSGS: Multi-Specification Robot Grasping System Based on Depth Camera

FU Lanhui1(),QIAO Weilai1,DENG Fuqin1,ZHOU Lei2,*(),HOU Jianfeng1,LENG Xiongwei3,ZHOU You2,HUANG Guangjun2   

  1. 1. School of Electronics and Information Engineering, Wuyi University, Jiangmen, Guangdong 529020, China
    2. Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, Guangdong 518132, China
    3. Ouotient Kinematics Machine, Dongguan, Guangdong 523000, China
  • Received:2025-03-02 Online:2026-02-20 Published:2026-02-02
  • Contact: ZHOU Lei

RichHTML

1

PDF

0

摘要：

【目的】机器视觉被广泛应用于工业场景下的物料识别与分拣等生产过程。当前的一些抓取方法存在抓取效率低、规格单一等问题。【方法】本研究提出一种多规格的机器人抓取系统MSGS（Multi-Specification Grasping System），通过优化YOLOv5模型对多种物料进行检测分类，根据不同型号物料的尺寸抓取系统可以自适应控制机械臂末端的伸缩距离，从而适应不同场景下多种尺寸规格的物料分拣任务。【结果】实验表明，检测模型能完成对小特征目标的检测与分类，准确率达到98.3%，mAP为0.993。通过多次进行定位与抓取实验，机械臂末端能根据物料规格伸缩末端距离并完成抓取行为，抓取成功率为100%，平均收缩误差为5 mm。【结论】该抓取系统能够较好地完成多种规格物料的分类、定位与抓取，为工业视觉的发展提供了一定的技术参考。

关键词: 机器视觉, 多规格, 目标检测, 无序抓取, 深度相机, 系统

Abstract:

[Objective] Machine vision is widely used in industrial scenarios for material recognition and sorting in production processes. However, some current grasping methods suffer from low efficiency and limited adaptability to single specifications. [Methods] This study proposes a Multi-Specification Grasping System (MSGS) for robots, which optimizes the YOLOv5 model to detect and classify various materials. The system can adaptively control the extension distance of the robotic arm's end-effector based on the size of different types of materials, thereby accommodating sorting tasks for materials of various sizes in different scenarios. [Results] Experiments show that the detection model can successfully detect and classify small-feature targets, achieving an accuracy rate of 98.3% and an mAP of 0.993. Through multiple positioning and grasping experiments, the robotic arm's end-effector can adjust its extension distance according to the material specifications and complete grasping actions with a success rate of 100% and an average contraction error of 5mm. [Conclusions] This grasping system can effectively perform classification, positioning, and grasping of materials with multiple specifications, providing valuable technical insights for the advancement of industrial vision.

Key words: machine vision, multiple specifications, target detection, random grasping, depth camera, system