Ontology-based 3D Pose Estimation for Autonomous Object Manipulation Rigas Kouskouridas, Theodora Retzepi, Eleni Charalampoglou and Antonios Gasteratos Democritus University of Thrace Department of Production and Management Engineering Vas. Sofias 12, Building I, Xanthi Greece 67100 Emails: rkouskou@pme.duth.gr tretzepi@ee.duth.gr echaral@ee.duth.gr gasteratos@ieee.org Abstract—In this paper a novel solution to the problem of guiding a robotic gripper in order to perform manipulation tasks, is presented. The proposed approach consists of two main modules corresponding to the training and testing sessions, respectively. During training, we employ an ontology-based framework with a view to the establishment of a database holding information regarding several geometrical attributes of the training objects. An accurate estimation of the 3D pose of an object-target is obtained during the testing phase and through the efficient exploitation of the established database. The most common solution to the 3D pose estimation problem implies extensive training sessions that are based on oversampled datasets containing several instances objects captured under varying view- points. However, such an approach engenders high complexity accompanied by large computational burden. We address this issue by proposing an ontology-based framework and a fuzzy- based approach that is able to efficiently interpolate between two known instances of the trained objects. Experimental results justify both our theoretical claims and our choice to adopt an ontology-based solution. I. I NTRODUCTION An expansive range of domestic tasks for service robots, such as collecting objects, loading or unloading a dishwasher and opening doors, is based on object handling. The problem of particle manipulation is specifically challenging in unstruc- tured environments that may include a range of several objects with varying shapes and sizes. Although humans are capable of excelling under such variations, a robotic application with the sufficient trade offs between performance and computational burden has yet to be built. In the last few decades, increas- ing importance has been gained to the problem of grasping unknown objects in a fully automatic way, mainly due to the wide-spread use of service and rehabilitation robotics [1], [2], [3], [4]. A psychological, biological and engineering focus has given to the manipulation task but is still considered as not being fully solved. Despite the existence of abundant available approaches for certain cases, there is still no general valid solution. According to the literature, the approaches dedicated to object manipulation are categorized into two major streams, engineering-based methodologies and vision-based strategies. Regarding the first category and given the importance of grasping for robots, a range of approaches have been proposed. Up to the last decade, most of these techniques relied on complete and accurate 3D models of the objects, in order mechanical operations accompanied by conventional methods to be employed. Building accurate models for an efficient representation of objects constitutes a very challenging task that often is sufficiently accomplished via laser scanning. A system for grasping 3D objects with unknown geometry using a Salisbury robotic hand was presented in [5],where each object was placed on a motorized and rotated table under a laser scanner in order a set of 3D points to be generated. These points were combined to be form a 3D model. A framework of automatic grasping of unknown objects via a laser-range scanner and a simulation environment was developed in [6]. Furthermore, a method for the adequate accomplishment of industrial bin picking tasks was presented in [7]. The authors proposed a system that provides accurate 3D models of objects that are further exploited in order to perform precise grasp- ing operations. However, the proposed super quadrics based object modeling approach can only be used for rotationally symmetric objects. Moreover, a technique to calculate possible grasping points for unknown objects with the help of the flat top surfaces of the objects based on a laser-range scanner system was published in [8]. Additionally, surface properties, such as friction and compliance are of basic importance in the grasping process. Nevertheless, a global metric cannot easily describe such attributes, whilst they are often modeled as being uniform for a whole object. An alternative approach to the object manipulation problem is the use of statistical learning methods. For instance, de Granville et al. [9] examined the problem of representing the orientation of a hand as it ap- proaches an object, and determined the feasibility of extracting canonical grasps from a human demonstration. Canonical grasps were represented using clustering procedures based on a combination of distributions. Another approach [10] involves combining analytical and empirical methods by segmenting an object into a set of super quadratics and then learning which ones are more suitable for grasping. According to the literature, an acceptable solution to the object manipulation problem could be given by integrating geometrical attributes of objects based on CAD models. Relatively simple CAD wire structures of objects are used by model-based methods [11], [12]. In this paper we present a new solution to the problem of training a robotic gripper in order to execute manipulation tasks. The suggested technique is composed of two units, namely the training and the testing modules. During training 978-1-4577-1775-8/12/$26.00 ©2012 IEEE