Deep Learning-Empowered Robot Vision for Efficient Robotic Grasp Detection and Defect Elimination in Industry 4.0

Abstract

Robot vision, enabled by deep learning breakthroughs, is gaining momentum in the industry 4.0 digitization process. The present investigation describes a robotic grasp detection application that makes use of a two-finger gripper and an RGB-D camera linked to a collaborative robot. The visual recognition system, which is integrated with edge computing units, conducts image recognition for faulty items and calculates the position of the robot arm. Identifying deformities in object photos, training and testing the images with a modified version of the You Only Look Once (YOLO) method, and establishing defect borders are all part of the process. Signals are subsequently sent to the robotic manipulator to remove the faulty components. The adopted technique used in this system is trained on custom data and has demonstrated a high accuracy and low latency performance as it reached a detection accuracy of 96% with 96.6% correct grasp accuracy.