Acquisition and Application of a Tactile Database Matthias Sch¨ opfer, Helge Ritter and Gunther Heidemann Abstract—We present a database of 2D pressure profile timeseries as a testbed for tactile object and surface recognition. The tactile database captures the surfaces of household and toy objects by moving a 2D pressure sensor mounted to an industrial robot arm around the objects using real-time trajectory calculation. Thus, it represents different “views” of the objects in a similar way as the well known Columbia Object Image Library (COIL) captures different views of an object by a camera. As a first application, objects in the database are classified using a neural network architecture. Index Terms— Robot tactile systems, Tactile systems (non- biological), Pattern Recognition, Image databases, Neural Net- works I. I NTRODUCTION Tactile sensing is a key discipline for exploration and grasping with autonomous robot hands. Humans can grasp and manipulate objects mostly without looking, guided only by haptics. But to date, there is no technical equivalent to human skin, in spite of research on different sensor designs [1], [2], [3], [4], [5], [6], [7]. So we still lack the “CCD- Chip” for tactile sensing. But the hardware problem is only part of the story. Even though suitable pressure sensors gradually become available, tactile recognition is still a rare research subject. Why is that so? One reason is that recognition of shape and surface structure without the help of the visual modality is a challenging pattern recognition task. However, the same is true e.g. for much better researched field of computer vision. In the opinion of the authors, a major obstacle in the way towards tactile pattern recognition is the complete lack of any standardized databases for realistic benchmarking of new algorithms. While in computer vision research databases such as the Columbia Object Image Library (COIL) [8] or VisTex [9] are widely established benchmarks, there is no such database for tactile sensing, because the gathering of tactile data is much more difficult than the acquisition of pictures, as will become clear in the following sections. In this paper, we present a database of tactile pressure profile timeseries which can be regarded as the “tactile equivalent” to COIL [10], [8]. COIL is a collection of images from 100 household and toy objects, where each object is represented by 72 different views, taken at angles 0, 5, 10 . . . degrees while the object rotates on a turntable. The database thus allows testing recognition algorithms This work was funded by the Deutsche Forschungsgesellschaft (DFG) Helge Ritter and Matthias Sch¨ opfer are with Faculty of Technology, Neuroinformatics Group, Bielefeld University, 33615 Bielefeld, Germany. {helge, mschoepf}@techfak.uni-bielefeld.de Gunther Heidemann is with Intelligent Systems Group, Stuttgart Univer- sity, D-70569 Stuttgart, Germany. . ais@vis.uni-stuttgart.de with various training and test sets for evaluation of view sensitivity and generalization. How can a “tactile COIL” be designed? First, we will have to decide on the sensor. If we want to acquire tactile data in a way similar to human sensing, a 2D sensor as a “fingertip” is required that is moved smoothly over the surface of an object. Basically, there are two ways to do this: The fingertip can either slide over the object, or “roll” without sliding. Since the real time control of sliding is highly difficult, we will roll the sensor over the object. Second, it has to be defined what we mean by a “view” of the object. While a view of an object is defined by the camera position relative to the object for image acquisition, there are much more degrees of freedom for the active acquisition of tactile data. Not only can the object be presented to the sensor in different positions and poses, but also the trajectory by which the sensor rolls over the object surface can vary. Therefore, the representation of each object in the database is characterized by two different sets of parameters: Object pose parameters and trajectory parameters. In this paper, we present a complete setup of a robot mounted tactile sensor, which is used to record haptic data (i.e. tactile and kinesthetic) from a set of 16 different small household and toy objects. Further, we describe a first application of the database: A neural recognition architecture which combines feature extraction by a local PCA approach with subsequent classification is used to analyze the high- dimensional tactile data for common patterns. We show, that classification of the object postures is possible and analyse the ability to generalize over different features. The reason for picking this application is not so much the demonstration of a certain use or practical application, but rather to show that the gathered data are valid, feasible and usable. A. Related Work The most active area in tactile sensing still seems to be the design and construction of new sensors [11]. Motoo, Arai and Yamada recently proposed a novel piezoelectric fingertip sensor [12]. Kim et. al. present a low cost 3 component tactile device array [13]. The soft tissue structure of the human finger for texture sensing is imitated in a design by Mukaibo et. al. [14]. Murakami and Hasegawa use a rather convetnional approach as they utilize a 6 DOF force torque sensor with a soft fingertip to detect edges and their direction [15]. All of these papers have in common, that they include some more or less extensive experimental part. Fewer papers or articles cope with algorithms on tactile data[11]. Platt, Fagg and Grupen propose a control basis for force-based interaction [16]. Several papers deal with the 2007 IEEE International Conference on Robotics and Automation Roma, Italy, 10-14 April 2007 WeE7.3 1-4244-0602-1/07/$20.00 ©2007 IEEE. 1517