Multiple-superquadrics based object surface estimation for grasping in service robotics

Abstract

A method for approximating an object's surface defined by 2.5D data is presented. The goal is to determine a compact volume which may be subjected to manipulation actions under a proper grasp planning. The object volume is obtained based on a union of parameterized geometric shapes, also known as superquadrics. In order to reduce over-segmentations and also the number of fitting combinations, the object space is defined as a 3D bounding Region of Interest (ROI). Further, via voxel decomposition, the ROI is decomposed in more meaningful smaller ROIs that can capture a large number of features from the object of interest. By using a fixed size voxel grid, the process can be slowed down for the cases of large objects. A boosting speedup of the process is proposed through dynamically adjusting the sides of the voxels for each object. Finally, a method for voxels merging is proposed. Each of the newly created ROIs, related to a particular region of the object, will be hosting a superquadric model which optimally estimates the considered surface.