Abstract
We demonstrate an inverted-pendulum mobile robot learning from observation of human motion. The robot is composed of two links, upper and lower bodies, two parallel wheels. It tries to imitate a kicking motion shown by a human. The imitation learning has three stages, that is, motion estimation, motion learning, and updating parameters of body mapping. The robot observes human demonstration with a camera and maps the silhouette of the human in the captured image to its own two links. It estimates of a link posture trajectory of the motion in which it tries to imitate the motion shown by the human. Then, the robot start to learn the kicking motion from the obtained link posture trajectories as initial motion parameters. Body mapping parameters affect performance of the imitation learning because initial motion parameter to be learned is sensitive to the performance. In order to obtain an appropriate body mapping parameter, learned motion is evaluated after the learning and the parameter is tuned based on the evaluation.
