Scholarlink

Pick-and-place (P&P) tasks in robotic applications have attracted considerable interest for several years. As these are repetitive tasks that are performed continuously and uninterruptedly, any slight improvement in the efficiency of this process results in a significant increase of productivity. In this work, we propose improving a pick-and-place task performed by two Dobot Magician educational manipulator robots alongside a conveyor belt. This system natively incorporates a hardware-based color sensor that allows the robot to identify and classify objects on a conveyor belt, organizing and palletizing them in specific columns according to their color characteristics. However, the robot spends time picking up the objects, placing them on the sensor, and then beginning palletization. In this work, we propose using the Faster R-CNN object detection model with ResNet-50 for feature extraction as an external and independent vision system to streamline and reduce the time losses associated with the hardware-based color sensor tasks. The implemented model achieved a mean Average Precision (mAP) of 68.0% during validation and 65.32% in real-world operation. The system accurately detects all key components involved in the task, including robotic arms, end effectors, the conveyor belt, the color sensor, and the colored blocks (red, green, and blue). When compared to the hardware-based color sensor method, the proposed approach reduced the total processing time from 249.31 s to 196.65 s, representing an improvement of 52.66 s, or approximately 21.12% in overall efficiency. This demonstrates the potential and promising results of using deep learning models in pick-and-place-related applications.

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

Implementation of Convolutional Neural Networks for Object Detection in Robotic Pick-and-Place Processes

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Renato Villazón,
  • Andrea Pilco,
  • Alvaro Javier Prado,
  • Viviana Moya,
  • Fernando Chicaiza,
  • Juan Pablo Vásconez

摘要

Pick-and-place (P&P) tasks in robotic applications have attracted considerable interest for several years. As these are repetitive tasks that are performed continuously and uninterruptedly, any slight improvement in the efficiency of this process results in a significant increase of productivity. In this work, we propose improving a pick-and-place task performed by two Dobot Magician educational manipulator robots alongside a conveyor belt. This system natively incorporates a hardware-based color sensor that allows the robot to identify and classify objects on a conveyor belt, organizing and palletizing them in specific columns according to their color characteristics. However, the robot spends time picking up the objects, placing them on the sensor, and then beginning palletization. In this work, we propose using the Faster R-CNN object detection model with ResNet-50 for feature extraction as an external and independent vision system to streamline and reduce the time losses associated with the hardware-based color sensor tasks. The implemented model achieved a mean Average Precision (mAP) of 68.0% during validation and 65.32% in real-world operation. The system accurately detects all key components involved in the task, including robotic arms, end effectors, the conveyor belt, the color sensor, and the colored blocks (red, green, and blue). When compared to the hardware-based color sensor method, the proposed approach reduced the total processing time from 249.31 s to 196.65 s, representing an improvement of 52.66 s, or approximately 21.12% in overall efficiency. This demonstrates the potential and promising results of using deep learning models in pick-and-place-related applications.