Real-Time Pose Imitation by Mid-Size Humanoid Robot With Servo-Cradle-Head RGB-D Vision System

Abstract

To begin with, the target human (TH) in the face of a mid-size humanoid robot performs 3-D motions captured by the servo-cradle-head RGB-D vision system (SCH-RGB-D-VS) on its head. During imitation processing, the SCH-RGB-D-VS can maintain a suitable field of view in the pitching and rolling directions to acquire the correct images of the TH's motion. Its necessity is first confirmed by the experimental result. Based on the 3-D coordinates of the head and two feet, 11 stable motions of the lower body (LB) are classified by the proposed improved support vector machine. Two pairs of hands and elbows for upper body (UB) imitation are approximated by eight pretrained multilayer neural network models to enhance one-to-one mapping, reduce the modeling complexity of inverse kinematics, and imitate complex motion. Finally, three categories of experiments by integrated motion of UB and LB confirm the effectiveness and robustness of the proposed method.