Neural network-assisted robotic vision system for industrial applications

Abstract

This paper investigates the performance of neural networks in positioning an industrial robot manipulator based on image feedback. Visual servoing regulates the pose of the robot manipulator in accordance with the target image data. As the servoing proceeds, the end-effector positions to its final pose as the image feature error exponentially decreases to zero. In this paper, the trained network moves the robot manipulator from an arbitrary initial pose to an intermediate posture based on the reference image features given. Then, fine-tuning based on the traditional visual servoing method is performed to achieve an accurate pick-and-place task with minimum iterations. The experimental results prove the capability of the neural network architectures to predict the desired pose using local image descriptor, corner points. This paper investigates the performance of two neural network designs applied to visual positioning and compares them with the traditional image-based visual servoing (IBVS) method in terms of execution time.