Prediction of object manipulation using tactile sensor information by a humanoid robot Shigeyuki Uematsu, Yuichi Kobayashi, Akinobu Shimizu and Toru Kaneko Abstract— This paper presents a framework of lifting-up manipulation acquisition based on tactile sensing informa- tion by a humanoid robot. Feature extraction from sensor information, including tactile information, is presented using linear and nonlinear mappings. Information acquired from sensors is mapped to a lower-dimensional space for predicting success of lifting-up task. Robot judges success or failure of the manipulation using the obtained feature space and object orientation. The proposed method was evaluated by simulation with a humanoid robot. Sensor information obtained at the beginning stage of lifting-up task was utilized to predict whether the robot can accomplish the task without dropping down the object. It was verified that the proposed feature extraction provides sufficient information to predict success of the task. The prediction will be utilized to modify posture of the robot. I. INTRODUCTION Nowadays, demands for robots that can perform tasks that have been conducted by humans are increasing [1]. One of them is to assist people who need nursing care to do their daily activities by themselves. Assistance robot for cooping meal [3] and a smart wheelchair [2] are examples of such applications. Another demand is to decrease heavy load handled by caregivers at nursing home and hospitals. Mukai et al. have developed a nursing-care assistance robot RIBA that can lift a human in its arms [5]. They used soft tactile sensors equipped at its arms for lifting up a bedridden patient. One of the difficulties of lift-up motion in such application is that object (patient) can be variant, in the sense of size, shape, posture, cloths, and so on. Thus, it is very important for such robots to cope with variance of object state. RIBA, as described above, has tactile sensors on its arm and has a large potential to adaptively manipulate object in various states, because tactile information provides rich information of contact between robots and the object [4]. Ohmura et al. realized whole-body contact motion to lift up a heavy object using tactile feedback [7]. Chitta et al. proposed a method for estimating state of objects using information obtained by tactile sensors [6]. One of the problems common to those approaches to effec- tively utilize tactile information is that most control strategies are problem-specific, designed suitable for each application. In conventional manipulation with a robot hand, stability of grasping has been formulated by analytical models, such as force closure and form closure [8]. Performance index of S. Uematsu is with the Department of Electrical and Electronic Engi- neering, Tokyo University of Agriculture and Technology, Koganei, Tokyo, 184-8588 Japan 50011645205@st.tuat.ac.jp manipulation was further extended to multi-robot manipu- lation case [9]. In the above-mentioned applications with humanoids, however, those idea cannot be directly applied due to difficulties of modeling contact between objects and robots. One possible solution for this problem is to introduce a data-driven approach. By directly observing many sample motions and their results, we might be able to evaluate a manipulation strategy from those observed experiences. If a robot can extract features that help to evaluate (or predict) a manipulation strategy autonomously, the method can be applied to cases with various objects and robots. In addition, such evaluation of a manipulation strategy could be equivalent to stability analysis of grasps without any analytical models. In this paper, a feature extraction method for evaluating manipulation of object by a humanoid robot is proposed. Evaluation is conducted through predicting whether current posture of manipulation is suitable for final achievement of manipulation task. Feature extraction based on mapping to lower-dimensional space and clustering is applied for estimation of success/failure of the lifting task. The proposed method is verified by simulation with a humanoid robot which has tactile sensors on its arm. The rest of the paper is organized as follows. Problem settings for the proposed method are described in section II. The proposed method of predicting task achievement is described in section III. The proposed method is evaluated in simulation in section IV. After discussing results of simulations in section V, section VI gives conclusion of the paper. II. PROBLEM SETTINGS A manipulation task of lifting-up an object by a humanoid robot is considered. An exemplar in simulation is shown in Fig.1, where configuration of humanoid robot NAO [13] is simulated by WEBOTS simulator [14]. There are two cameras at the head of the robot, but both of them do not provide sufficient information about relative position of an object when it is located near to the robot. Measurement of object close to the robot by visual information is especially difficult in cases with a large object with few textures, because viewing range will be filled with the object without any visual cue to detect its position. It is assumed that force sensors are attached on each arm of the robot so that it can evaluate contact with an object * . * The original configuration of NAO does not include force sensors. 978-1-4673-2706-0/12/$31.00 ©2012 IEEE