Real-time imitation framework for humanoid robots based on posture classification

摘要

A safe real-time imitation framework for humanoid robots to human beings action is proposed. Postures of an actor are divided into three categories: robust postures, normal postures and dangerous postures. To deal with the robust postures, direct mapping from skeleton data gathered by 3D cameras to each angle of robot joint motors is taken because the postures are robust enough for imitation in high speed. When it comes to a normal posture, one specific safe robot posture is designed to correspond to each part of separated human postures. As the actor changes his posture from one to another, the robot will also switch from the corresponding posture to another in the designed way. Meanwhile, when there is no need to switch, the direct mapping is taken to fulfill the task. The imitation strategy for dangerous postures is the same as the normal postures except that the robot can not use direct mapping when there is no switch task because some changes of the designed posture may lead to tumble.