ConvNets for 3D Coordinate Prediction and Direct Visuoposition Control of a Robot Arm

• Main Author: 張育鳴
• Other Authors: LI ,CHIH-HUNG
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/ym3gfp

碩士 === 國立臺北科技大學 === 製造科技研究所 === 107 === In this study, as part of our effort in SLAM research, the goal is to design a precision robot control algorithm, based on convolutional neural network and image processing. This study creates a real-time visual classifier and detector that can be automatically generated in a short time to locate the target object for intelligent control of the robot arm. The proposed architecture only needs to capture a single base environment image. Through the structured processing of the training image, the convolutional neural network can be used to predict the three-dimensional coordinates of the target object. We call this method ONE SHOT, which can be applied to the arm control on the mobile robot and can be used to perform the last action of the arm, such as pressing the elevator button. Through the double-point positioning of the double photography, the angle correction can be quickly performed to achieve the planar positioning of large objects. The method can be used for picking and placing of the workpiece in the automation equipment. Our architecture adopts a typical PC and can perform real-time control at an image detection rate of 50fps. The experimental results confirm that the advantages of this study are (a) short preparation time, the training set required for convolutional neural network training is automatically generated for the base image without manual annotation; (b) high flexibility and robustness, for the environment, the high adaptability is hundreds of times faster than the template matching in the detection speed; (c) the precision is high, and the motion positioning of the robot arm can achieve the precision of 0.5 mm.