Robot end-effector recognition using modular neural network for autonomous control

Abstract

Control systems using a computer vision technology can be very useful to control and track movements of a robot at a remote site. For this purpose, we propose a robot end-effector recognition system which automatically adapts to given conditions using a modular neural network. The robot end-effector can be seen in various shapes in images including translation, rotation and scaling of an original end-effector image. The designed modular network consists of several modules of neural networks, each of which employs a multilayer feedforward neural network. The modular neural network is used to recognize the robot's end-effector precisely and to minimize the processing time. Our experiments show that the performance of the modular neural network is superior to that of the single module neural network. The control system using the suggested modular neural network can perform the recognition with a high precision even under the very hard conditions : occlusion, cluttering, or various other motions.