Camera-space manipulation with natural visual information

Abstract

Camera-space manipulation is a precise and robust approach to robot hand/eye coordination in practical, three-dimensional situations. The acquisition of visual information necessary for camera-space manipulation has so far been simplified by placing easily detectable visual cues on the objects involved in the task. This work concerns the development of an approach to speedy and precise extraction of visual information from features on the objects. This approach invokes a symbiosis between the estimation procedure of camera-space manipulation and image analysis, in which each procedure builds on progressively improving information learned by the other, both during a particular cycle of a repetitive task, and from one task-cycle to the next. The idea was implemented in a three-dimensional automobile wheel-loading task, in which the natural features were circular. The relative locations of video cameras were such that error-propagation from the estimation procedure to object positioning would be near a minimum. The image-analysis procedure developed in this work could respond to the availability of better information. Its speed and precision improved continually with improving accuracy of the estimation procedure. This procedure can be expected to be more suitable for camera-space manipulation than methods based on the Hough transform. Finally, a method to generalize the approach to tasks involving a class of features was proposed and tested.