ABSTRACT

This thesis describes a novel approach to realizing self-learning calibration-free stereo-vision-guided robots. It makes no use of world coordinates and of inverse perspective or kinematic transformations. Thus, it eliminates the need for a calibration of the robot and of the vision system and adapts automatically to changes of the system parameters and of the environment. Also, it allows the robot to acquire and accumulate automatically whatever knowledge it may need into its knowledge base and to gain experience in the course of its normal operation and, therefore, to improve its skills and operating speed over time. Thus, the robot presents characteristics of a self-learning system.

Key points of this approach are the independence of any special infrastructures, the absence of any quantitative knowledge of the mechanical and kinematic characteristics of the robot and of the visual system, as well as of the objects to be grasped and the renunciation of a world coordinate system; the control of the robot is accomplished by a direct transformation of sensor data, e.g., image coordinates, to motor control commands. As a result of this, the robot displays a remarkable degree of robustness against parameter changes.

The objects to be grasped are detected from the background and other irrelevant image information based on generic 2-D object models of the object appearance in the images. The object pose, i.e., position and orientation, and the suitable grasping points are automatically determined according to the object appearance recognized in the stereo images.

The approach has been implemented on a calibration-free vision-guided manipulator with five degrees of freedom (DOF) and was evaluated in real-word experiments. Depending on the shape and orientation of the object to be grasped, all five degrees of freedom may be controlled and used. The experiments have shown that grasping of a variety of differently shaped objects that may initially be located in arbitrary positions in the robot's 3-D work space, and automatic skill improvement are, indeed, possible.

Keywords:      Calibration-free Robots, Intelligent Robots, Knowledge-based Systems, Learning Robots, Machine Learning, Object Grasping, Object Recognition, Robot Vision, Situation-oriented and Behavior-based Stereo Vision, Real-time Image Processing, Vision-guided Robots.