ロボティクス Robotics

P2-1-249 ヒューマノイドロボットのモーションデータにおける運動プリミティブの自己組織化 Self-organization of motor primitives from motion data of humanoid robot ○久保田守1, 山崎匡2, 西野哲朗1 ○Mamoru Kubota1, Tadashi Yamazaki2, Tetsuro Nishino1 電気通信大学大学院 情報理工学研究科 総合情報学専攻1, 電気通信大学 大学院 情報理工学研究科 情報・通信工学専攻2 Graduate School of Informatics and Enginerring, The University of Electro-Communications, Tokyo1, Graduate School of Informatics and Enginerring, The University of Electro-Communications, Tokyo2 Hobby-use humanoid robots are small and inexpensive toy robots typically composed of about 20 servomotors for joints. Dynamic motion of these humanoid robots is defined by a sequence of static poses. Each pose is defined by specifying joint angles numerically for each joint.This is a difficult task. First, defining the exact angle values for more than 20 joints for each pose is troublesome. Second, mixing more than two motions to create a new one is difficult. Third, when we need a motion similar to an existing one with a slight change, we have to define a new sequence completely. All of them come from the fact that the programming is based on joint angles. We need a way to represent robot motions abstractly.Our brain employs a different strategy for motion generation. Instead of specifying each joint angle or each muscle tone, the primary motor cortex has a set of primitive motions called motor primitives, and higher cortical areas and the basal ganglia generates a sequence of combinations of such motor primitives. Thus, our brain uses more abstract form to compose dynamic motions.In this study, we propose a method to represent a motion of a hobby-use humanoid robot by combination of the motor primitiveswhich acquired by self-organizing map (SOM). We hypothesize that the primitives are acquired through learning of a number of motion examples in self-organizing manner, because motor primitives should be statistically meaningful components that are sufficient to organize most of daily movements by combination. To justify it, we employ a robot and create some motions by hand. Then, we adopt statistical learning algorithms for self-organization to extract motor primitives from the motions. Finally, we demonstrate that the original motions are reproduced by the combination of the extracted motor primitives. These results suggest that motion representation based on motor primitives provides a more abstract and efficient way to make motion for hobby-use humanoid robots.

P2-1-250 ユーザ適応型支援ロボットはダーツ投げ動作学習を促進する User-Adaptive Robotic Training Accelerates Learning of Darts Throwing ○大林千尋1, 為井智也1, 柴田智広1 ○Chihiro Obayashi1, Tomoya Tamei1, Tomohiro Shibata1 奈良先端科学技術大学院大学　情報科学研究科1 Graduate School of Information Science, NAIST, Nara1 Acquiring skillful movements of experts is a difficult task in many fields. One of its difficulties is represented by guidance hypothesis; humans tend to rely too much on external assistive feedback, resulting in interference with internal feedback necessary for motor skill learning. A few robotic training studies have attempted to deal with this difficulty, but they required a physical model of the user's motor system which is essentially difficult. They also required a predetermined desired trajectory to be followed by learners, but it is unclear whether it is optimal for each learner. Here we show that our robotic training system that is user-adaptive and that does not require a desired trajectory nor the physical model of the user's motor system can accelerate learning darts throwing of novices. The adaptability was provided by a policy-gradient reinforcement learning algorithm which is a model-free type, and whose goal was maximizing the score of the darts throwing and minimizing physical support. A soft physical support for the upper arm was provided from underneath by a robot manipulator because we found that there was a significant difference in the elbow-position stability between novices and experts. The manipulator that was impedance-controlled in the vertical direction, and its spring parameter was changed during training according to the learning algorithm. The manipulator was not mechanically coupled with the upper arm so that the system does not impose a desired trajectory as much as possible. 18 novices were participated in our training experiments over two days; they were randomly divided into three groups having different experimental conditions: (1) without-robot, (2) with non-adaptive-robot, and (3) with-adaptive-robot. As a result, condition (3) showed the significant increase of the score compared with the other conditions (Tukey-Kramer, p<0.05), suggesting the plausibility of the proposed training system.

上部に戻る

前に戻る