Learning Continuous Grasp Stability for a Humanoid Robot Hand Based on Tactile Sensing J. Schill and J. Laaksonen and M. Przybylski and V. Kyrki and T. Asfour and R. Dillmann Abstract— Grasp stability estimation with complex robots in environments with uncertainty is a major research challenge. Analytical measures such as force closure based grasp quality metrics are often impractical because tactile sensors are unable to measure contacts accurately enough especially in soft contact cases. Recently, an alternative approach of learning the stability based on examples has been proposed. Current approaches of stability learning analyze the tactile sensor readings only at the end of the grasp attempt, which makes them somewhat time consuming, because the grasp can be stable already earlier. In this paper, we propose an approach for grasp stability learning, which estimates the stability continuously during the grasp attempt. The approach is based on temporal filtering of a support vector machine classifier output. Experimental evaluation is performed on an anthropomorphic ARMAR-IIIb. The results demonstrate that the continuous estimation provides equal performance to the earlier approaches while reducing the time to reach a stable grasp significantly. Moreover, the results demonstrate for the first time that the learning based stability estimation can be used with a flexible, pneumatically actuated hand, in contrast to the rigid hands used in earlier works. I. INTRODUCTION The sense of touch is essential to human grasping. The work described in this paper considers robotic tactile sense as a biomimetic replacement for the sense of touch, especially when estimating grasp stability. Grasp stability in analyti- cal sense is well defined and can be readily computed in simulation where enough data of the grasp is available, i.e. all contacts between the robotic hand and the object that is grasped. Additionally, using a force closure metric for grasp stability, one can compute a grasp that sufficiently resists outside forces, such as gravity, thus allowing the robot to manipulate the object, for example by lifting the object. However, when using real hardware, the tactile sensor data is imperfect, both in the sense of detecting contacts and in the sense of determining the actual contact forces. In some cases the proprioceptive information, i.e. joint configuration, is also difficult to determine accurately, thus, causing uncertainty in ascertaining the kinematic configuration of the hand. All these described phenomena pave a difficult road for computing the grasp stability analytically with real hands. The research leading to these results has received funding from the European Community’s Seventh Framework Programme GRASP under grant agreement n ◦ 215821 and Xperience under grant agreement n ◦ 270273 J. Schill, M. Przybylski, T. Asfour and R. Dillmann are with the Humanoids and Intelligence Systems Lab, Karlsruhe Institute of Tech- nology, Karlsruhe, Germany, {markus.przybylski, schill, asfour, dillmann}@kit.edu J. Laaksonen and V. Kyrki are with Department of Information Technol- ogy, Lappeenranta University of Technology, P.O. Box 20, 53851 Lappeen- ranta, Finland, jalaakso@lut.fi, kyrki@lut.fi In this paper, we focus on learning the grasp stability instead of analytically solving it. Compared to the analytical methods, learning requires training data, which needs to be collected beforehand. As the training data, we can use any pertinent data that can be collected from robotic hand, in our case we use input from all tactile sensors and the hand finger configuration. It is also important to notice that the raw sensor data can be used in the learning, for example, there is no need to know the kinematic configuration of the hand to compute the true locations of the contacts when analytically solving the grasp stability. This feature allows grasp stability to be learned for many different robotic hands with only minimum changes. There has been a number of publications on learning the grasp stability [1], [2]. These approaches evaluate the stability after the hand finished closing around an object. We extend the work presented in previous papers, so that the decision on the grasp stability can be achieved during the grasping instead of at the end of the grasp. We also demon- strate that the learning of the grasp stability is possible with the ARMAR-IIIb hand [3], [4], a flexible anthropomorphic hand operating on pneumatics. The rest of the paper is divided into four sections. Sec- tion II gives an overview on learning grasp stability as well as other learning approaches that are grasp and manipulation related. Section III introduces a theoretical background on machine learning methods and how they can be applied to the grasp stability problem. Section IV contains the experiments made on data collected using the ARMAR-IIIb hand. We conclude with discussion of the results in Section V. II. RELATED WORK Grasp stability analysis by analytical means is a well es- tablished field. However, to analytically determine the grasp stability, the kinematic configuration of the hand and the contacts between the hand and the object must be perfectly known. This subject has been well studied and [5] gives a detailed review on the subject. However, the references are useful only in cases when conditions described above are true. When this is the case, it is possible to determine if the grasp is either force or form closure grasp [6], which ensures the stability. Compared to this body of work, we wish to learn the stability from existing data, i.e. the tactile data. The research on use of tactile and other sensors in a grasping context has increased in last few years. Felip and Morales [7] developed a robust grasp primitive, which tries to find a suitable grasp for an unknown object after a few initial The Fourth IEEE RAS/EMBS International Conference on Biomedical Robotics and Biomechatronics Roma, Italy. June 24-27, 2012 978-1-4577-1198-5/12/$26.00 ©2012 IEEE 1901