No abstract

No abstract

No abstract

This thesis proposes a framework that enables a biped humanoid robot to learn and reproduce whole body motions from human motions. This ability can increase the usefulness of biped humanoid robots, which have a potential to perform human-like motions of the whole body. We achieve this goal on the basis of the Learning from Observation (LFO) paradigm, and prove that the paradigm is valid for whole body motions. In our framework, the target motions are dances, and a robot uses its own legs to support its body during a dance performance. Reproducing such motions from human motions is a novel attempt. This thesis especially focuses on leg motions, which present a serious challenge to the achievement of our goal.LFO is a paradigm of teaching a task to a robot. It consists of three processes: observation, recognition, and execution. First, a human instructor demonstrates a task and a robot observes the demonstration. Then the robot recognizes what is done in the observed motion. After that, the robot can execute the task according to the recognition results. In this paradigm, all an instructor has to do is to demonstrate a task with his or her body. This makes teaching simple and efficient because complex operations or programming by a robotic specialist is not required. This simultaneously means that a robot imitates an instructor's way of doing a task. These features are great advantages in teaching human-like motions to a robot that has a complex humanlike body.In the LFO paradigm, the recognition and execution processes are based on task models designed for a target task domain. Task models provide essential representations of tasks, which are necessary for executing tasks in various situations. One of the fundamental factors in different situations is the differences among bodies. In this study, to solve the problem caused by body difference is the most important role of task models.In general, human or humanoid bodies have their own physical characteristics different from each other. The differences include geometry of body parts, weight distribution, and physical capabilities. In addition, humanoid robots can have their own body mechanisms quite different from those of another body. Due to these differences, observed motions of an instructor cannot be performed directly by another body of a robot. It is hence inevitable that the motions are partly modified so that they are adapted to the body that executes the motions. This adaptation must be processed so that the original purpose of a task is achieved. For motions of the whole body including legs, body differences significantly affect the execution of motions because the motions are performed under severe constraints caused by the interaction with the floor. For motions such as dances, important characteristics of the original motions must be preserved simultaneously. These factors make reproducing whole body motions performed by another body a difficult problem. Since legs play a major role in the i interaction with the floor, leg motion is especially...