Visual servo based on CAD model using machine learning

Abstract

Image Based Visual Servo(IBVS) is known as a controlled method using data from vision sensors, which is to place the object into a specified place with a desired orientation. In order to place an object, the derivation of a matrix called image jacobian is necessary. The image jacobian can be used only for the image captured in the vicinity of the target image. There for, the rederivation of image jacobian is necessary when placing an object into different target places which takes time and efforts. In this paper, we propose a new visual servo based on deep learning and evaluate it in a simulation. By including various goal images in the data set, we achieved placing an object into different targets using a single trained convolutional neural network. Details of the data set and the architecture of convolutional neural network is described in this paper.