An anthropomorphic binocular-vision planning for grasping parts by robots

Abstract

Sensor planning by active vision having anthropomorphic features such as binocularity, foveas, and gaze control is proposed. The objective of the vision is to provide robots with the pose information of an adequate object to be grasped by the robots. For this, the paper describes the following: a viewer-oriented fixation point frame and its calibration; active motion and gaze control of the sensor to focus on a region or an object of interest; selection of a specific candidate object; disparity filtering; zoom control for variation from coarse to fine resolution; and the estimation of grasping pose of a specific portion of the object. On the basis of the importance of the contour information and the scheme of stereo vision in recognizing objects by humans, the occluding contour pairs of objects are used as inputs, in order to show the proposed vision planning.