The sensor-control Jacobian as a basis for controlling calibration-free robots

Abstract

A method for controlling the motions of robots is presented. It is based on the newly introduced sensor-control Jacobian matrix and avoids all quantitative modeling of the robot and the sensor system. The sensor-control Jacobian contains the coefficients that relate those changes in sensor data which are caused by a motion of the robot to the robot control words that caused the robot to move and, thus, the sensor data to change. A wide variety of tasks of robots can be reduced to minimizing the differences between actual sensor data and a set of hypothetical sensor data corresponding to some desired state. All these tasks can be solved by this method. The method is especially useful for calibration-free robots, since neither quantitative models of the mechanical, kinematic and control characteristics of the robot, nor knowledge of the sensor characteristics are required. The sensor-control Jacobian may be determined automatically in real time while the robot is operating. This yields a high degree of adaptability and flexibility against unforeseen changes in the robot's parameters. Because the concept has an open structure it allows further extensions and improvements, e.g., in terms of the utilization of sensor data redundancy and machine learning. For the purpose of evaluation, the concept has been implemented on a calibration-free camera-manipulator system. Real-world grasping experiments have demonstrated the effectiveness of the method.