Cooperation and competitive learning of multi-humanoid based on mapping other's behavior on the self behavior space

近年来，类人生物的发展引发了各种各样的研究问题。然而，在大多数研究中使用的控制是由设计者明确指定的。为了克服自顶向下设计的局限性，本研究项目旨在从模仿发展过程的角度出发，发展多类人自适应行为的自主学习方法。这些项目涉及四个主题。在第一个“身体表征”中，我们提出了通过行为经验获得感觉-运动映射的身体表征学习模型。利用获得的映射，实现了两个类人机器人之间的传球行为。在第二篇“行为的识别与模仿”中，我们提出了在姿态、动作和行为层面上自我与他人行为映射的学习模型。在获得映射之后，人类和一个更像人类的机器人之间的手势交流就被学习了。在第三章“他人意图的估计”中，我们提出了一种基于状态值估计的其他智能体意图推理的新方法。该方法不需要精确的世界模型或协调转换系统来处理视图依赖关系。该研究表明，观察者可以通过观察行为过程中估计的状态值转移来推断他人的意图，而不是通过笛卡尔空间中精确的物体轨迹。在力一“词汇习得”中，我们提出了一个词汇习得模型，该模型利用显著性将观察对象的视觉特征与看护者所发出的标签联系起来。机器人根据显著性改变其注意力和学习率。仿真实验表明，具有显著性的学习模型能有效地将给定的标签与观察到的特征相关联。此外，在真实的仿人机器人实验中，用独立于视点的自组织地图自适应地表示被观察物体的形状来表示视觉特征。少

Recent progresses in developing humanoids have opened various kinds of research issues. However, the controls used in most of the studies are specified explicitly by designers. To overcome the limitation of the top-down design, this research project aims at developing the methodology of autonomous learning of multi-humanoids for adaptive behaviors from the view point of the development process of imitation. The projects has dealt with four topics. In the first one, that is "body representation", we proposed the learning model for the body representation in which the sensori-motor mappings are acquired through behavior experiences. Using the acquired mapping, the ball pass behaviors between two humanoids are realized. In the second one, "Recognition and Imitation of behavior", we proposed the learning model for mapping between the self and the other's behavior in posture, motion and behavior levels. After acquiring the mapping, the gesture communication is learned between a human and a … More humanoid. In the third one, "Estimation of other's intention", we proposed a novel method of inference of other agent's intention based on the state value estimation. The method does not need a precise world model or coordination transformation system to deal with view dependency. This study shows an observer can infer an intention of other not by precise object trajectory in Cartesian space but by the estimated state value transition during the observed behavior. In the force one, "Lexical acquisition", we proposed a lexical acquisition model which makes use of saliency to associate visual features of observed objects with the labels uttered by a caregiver. A robot changes its attention and learning rate based on the saliency. Simulation experiments showed that the learning model with saliency effectively associate the given labels with the observed features. Moreover, in the experiment with a real humanoid robot, the visual features are represented with self organizing maps which adaptively represents the shape of observed objects independent of the viewpoints. Less