Generating Image Trajectories and Direct Image Jacobians for Visual Servoing

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

This is a three year standard award. Visual servoing, the use of visual feedback in the control of a robot end-effector, can provide a superior robot interface than those offered by teach pendants or tele-operated systems. Visual servoing is advantageous when precise fixturing is impractical or impossible. In image based servoing, control is effected in the image. This method has the advantages that one does not have to reconstruct 3D information about the robot's environment, and that control laws exist that are not overly sensitive to camera calibration errors. However, there are problems with image based servoing that need to be overcome in order to develop robust, stable systems. This project will address two issues impeding the application of image-based servoing. First, image feature-point path planning methods will be developed and demonstrated. Previous research efforts have focused on the form, and possible decompositions, of the "Image Jacobian", relating manipulator velocities to image motion. In contrast, this project will develop methods to determining a path within the image to use as input to a "standard" image-based regulator. The PI's approach will both improve controller performance and also ensure the image features remain within the field of view during the manipulation task. Second, a parameterized direct image-to-actuator model will be analyzed, evaluated, tested, and compared with classical approaches. The error signal from the image together with the image Jacobian typically map to end-effector velocities which then must be converted to actuator velocities using the robot kinematic model. A controller that converts image displacements directly to actuator commands would eliminate the intermediate step thus enhancing system performance, both in terms of speed and accuracy. Addressing issues of robustness and sensitivity will facilitate application of these techniques outside of the laboratory. Success of this project will facilitate the acceptance of image based servo techniques as a viable control method in production operations. The key scientific contributions created by this effort will include advances in the design of image based visual servo controllers through (1) the introduction of superior robot motion command generation created through image path planning and (2) the introduction of a parameterized direct image Jacobian with appropriate understanding of the model's limitations.