Real-Time Posture Imitation of Biped Humanoid Robot Based on Particle Filter with Simple Joint Control for Standing Stabilization

摘要

The purpose of this study is a development of real-time imitation learning system for a humanoid robot. It is needed to estimate joint angles of the robot from the observation of human demonstration, however, it is difficult to measure the joint angles directly. Conventional motion capture systems which measure the joint angles of a human subject precisely are often expensive or hard to use in daily life. Recently, depth sensor has become popular to provide fully-body 3D motion capture because it has advantages of low cost and no markers or trackers. It provides the position of joints of the human subject, however, the joint angles for the imitating robot have to be calculated in some way. Inverse kinematics is often used for the joint angle calculation, however, it needs relatively high computational cost for the optimization calculation and sometimes it has a difficulty to have a unique solution because of redundancy. We propose to use a particle filter for joint angle imitation so it provides a reasonable solution with a less computational cost to realize a real-time imitation of a humanoid robot through observation of human demonstration. However, the particle filter does not provide the standing stability of the humanoid robot. Therefore, we propose a novel and simple method of control of leg joints for the standing stabilization. While the humanoid robot imitates the human posture, the ankle and the hip joint angles of the robot are controlled based on the knee and hip joints provided by the particle filter. We evaluate the proposed method with experiments using a humanoid robot, Aldebaran Robotics NAO, and show its validity.