Vision-Based Perception of Articulated Objects

Abstract

The probabilistic framework developed in the previous chapter enables a manipulation robot to learn accurate kinematic models of articulated objects. As input, our framework requires a sequence of pose observations of the articulated object. We implemented the perception in the previous chapter using visual markers or by directly recording the end effector trajectory while the robot was manipulating the articulated object. For the daily use in domestic environments, however, both options are not satisfactory: clearly, it is neither desirable to augment all furniture with visual markers nor to guide a robot manually to the handles of all relevant objects.