Robust object localization based on error patterns learning for dexterous mobile manipulation

In this article we describe an approach for object detection and pose estimation from stereo RGB frames for robot manipulation in manufacturing scenarios. This solution was developed in the framework of the second challenge of the EuRoC project, and meets the need of a registration method invariant to the view perspective and robust to the structural symmetries and ambiguities of the target objects. Our contribution consists of automatic correction of sub-optimal results of registration algorithms. As most registration algorithms only converge on local optima, a tool for recognizing and correcting wrong alignments is highly desirable. Our insight is that, for a given target point cloud, it is important to study the alignment space offline and identify sub-optimal solutions before the registration. The convergence of the algorithm leads to the error pattern knowledge that can be used to discard the wrong solutions, and recover the correct alignment. Experiments on synthesized and real data show that exploiting the known information about the spatial properties of the objects, together with appropriate pre-processing and refining of the data, we can have a substantial improvement in discarding wrong hypothesis for geometrically ambiguous items.