Balancing and Reconstruction of Segmented Postures for Humanoid Robots in Imitation of Motion

This paper introduces an imitation system based on the similarity of the replaying motions of robots with the sequential poses of a demonstrator. The system is composed of three modules—key pose elicitation, real robot balance control, and memorization for motion replay. The elicitation of key poses drives the balance learning and motion replay of the robot. Dissimilarity values associated with the defined spatiotemporal function of simultaneous joint motion are used to analyze the degree of similarity. To overcome the difference in mechanical structures and kinematics, such as the number of joints between robots and human demonstrators, the key poses extracted from the motions of demonstrators are modified by a Q-Learning process that considers the kinematic constraints and maintains the balance of the robot while executing imitation. The rewards were designed not only to encourage a robot to execute as many consecutive poses as possible, but also to guide the robot to maintain its balance even though the biped lacks information on the ankle joint. These modified key poses are stored in databases for replaying or composing new motions in an ordered sequence. The experimental results demonstrate that a robot could adjust the poses, mapped from the movements of the demonstrator, to its static stable states, thereby imitating human motions by self-learning.