Motion replaying by humanoid robots based on segmented and stable postures

The proposed research focuses on how humanoid robots replay motions, which were demonstrated by human demonstrator, steadily. First, the proposed system captures the three-dimensional coordinates of a human body's joints by using Microsoft Kinect v2. Then, the captured joint-information is converted to the angles of biped's motors. Further, extract the key poses from these captured poses in order to reduce the quantity of joint information for a series of pose in a motion. For each key pose, use Q-Learning to adjust the angles of a biped's motors such that the biped can replay the captured pose steadily. Besides, a key pose database, which contains the balanced key poses, and a motion sequence database, which contains the keypose sequence of demonstrators' motions, are maintained for motion replaying. Finally, search motion sequence database, based on Sequence-Pattern-Mining approach, to find the most similar motion sequence to the motion, which the biped will imitate currently. The simulated results show that the biped could replay the demonstrators' motions effectively and steadily.