Tactile Sensing: Steps to Artificial Somatosensory Maps Giorgio Cannata, Simone Denei, Fulvio Mastrogiovanni Abstract— In this paper a framework for representing tactile information in robots is discussed. Control models exploiting tactile sensing are fundamental in social Human-Robot in- teraction tasks. Difficulties arising in rendering the sense of touch in robots are at different levels: both representation and computational issues must be considered. A layered system is proposed, which is inspired from tactile sensing in humans for building artificial somatosensory maps in robots. Experiments in simulation are used to validate the approach. I. I NTRODUCTION Robots exploiting tactile information are expected to ex- hibit advanced capabilities in physical and social Human- Robot Interaction (HRI in short). The sense of touch is a fundamental feature for control models based on physical interaction cues. Appropriate social and physical stimuli are needed to enhance the quality of the interaction in terms of robot behaviour and responsiveness. To date, studies in HRI have been largely devoted to inves- tigate suitable models for modulating interaction behaviours at the social level [17]. Aspects related to physical interaction received considerable attention mostly with respect to tactile sensing, and specifically to transduction technologies [5]. Although the need arises to integrate information from both physically and socially oriented models of interaction, the direct use of tactile data in designing control strategies enforcing social interaction rules did not receive considerable attention in literature. One possibility is to design appropriate representation structures to mimic somatosensory mapping in humans. From one side, these structure must guarantee an unique mapping between the tactile elements on the robot surface and their representation; from the other side, they must be accessible from high level behaviours implementing social models of interaction. During the past few years, a number of approaches have partially addressed these key issues. A model aimed at translating contact phenomena into language like symbols has been presented in [23], where the focus is more on the relationship between numerical and symbolic data rather than in the use of such information at the control level. The work presented in [15] faces the problem of the emergent behaviour through sensory-motor interaction between an agent capable of full body movements and the surrounding environment. A somatosensory map is obtained All the authors are with the Department of Communication, Computer and System Sciences, University of Genova, Via Opera Pia 13, 16145, Genova, Italy. Corresponding Author email:fulvio@dist.unige.it. The research leading to these results has received funding from the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement n. 231500/ROBOSKIN. by correlating signals from tactile sensors distributed over the agent surface. A similar approach, based on Informa- tion Theory, has been proposed in [20], where sensoritopic maps of groups of sensors are created using self-organizing processes. Feedback from groups of tactile sensors has been used in [19] to determine sensoritopic connections between correlated taxels: a mostly manual learning process is used to activate groups of nearby taxels, which are then considered topographically close to each other. Both representation and computational issues must be considered when designing artificial somatosensory maps: • Designing internal models for representing tactile in- formation is a novel research attempt. Differently from vision images, tactile images can not be easily flattened on a 2D metric space, since they originate from elements that are located on curved surfaces. Such concepts as proximity, feature extraction or data filtering can not be easily applied. Furthermore, cameras provide infor- mation from a well-defined location in space, whereas taxels are distributed over large parts of the robot surface, which are subject to kinematics constraints. A natural representation preserving skin topology must be available. • Data structures must guarantee an easy access to a semantically well-defined tactile information. Tactile data at different resolution must be accessible according to the task at hand: high resolution tactile images are needed for fine contact dynamics, whereas reactive be- haviours can be attained by manipulating low resolution information with an associated negligible computational load. The characteristics of the contact must be acces- sible from the representation. Approaches in the literature do not maintain the map- ping between the location of taxels in 3D space and their representation in the artificial somatosensory maps. This is fundamental to use tactile data in practice, e.g., when reacting to sudden contacts. The main contribution of this work is a model of somatosensory maps addressing representation issues, which is loosely inspired by tactile rendering in humans. However, it also paves the way for further develop- ments on the computational side, which are outside the scope of the present discussion. Inspired by the beautiful images of [24], a hierarchical architecture for tactile rendering is proposed that exploits Surface Parameterization techniques to model somatosensory maps. The paper is organized as follows. Section II describes the actual tactile information processing architecture in humans. 19th IEEE International Symposium on Robot and Human Interactive Communication Principe di Piemonte - Viareggio, Italy, Sept. 12-15, 2010 978-1-4244-7989-4/10/$26.00 ©2010 IEEE 576 978-1-4244-7990-0/10/$26.00 ©2010 IEEE