Development of a calibration-free vision based robot control system

Abstract

The author worked on using calibration-free visual feedback from camera sensors that were not calibrated and based on this visual feedback data obtained, a neural network was developed to approximate the Jacobian matrix. This was then implemented into a standard PD controller to control a robotic arm to perform setpoint task. Not having to calibrate the kinematics of a robot or of their sensors allows for lower maintenance cost and greater reliability. Otherwise, any changes in the camera positions or the robotic arm kinematic parameters will bring about the need for recalculation of perspective projection parameters which will be an expensive task. Aging of the robotic arm or any other maintenance work is tolerated without compromising on the robotic arm performance. The approach used is mounting of an uncalibrated camera above the robotic arm which will observe the position of the end effector and then the respective desired position coordinates(x’, y’) will be collected and saved to be used for the offline training of the neural network. The neural network algorithm developed approximates the Jacobian matrix which is then integrated into the robot system to reach a desired setpoint. The end effector of the robotic arm will gradually move to the desired setpoint specified by a user by using the approximated Jacobian matrix.