Pose Analysis of Humanoid Robot Imitation Process Based on Improved MLP Network

This paper proposes a system, by which humanoid robots could imitate human movements to learn human behaviors. Considering the security problem during humanoid robot imitation process, a method of stability identification based on improved Multi-Layer Perception (MLP) is proposed. Firstly, the positions of the human joints are collected by a Kinect. Secondly, the human joints information is transformed into robot angles information by the inverse kinematics equation. Finally, the stability of the robot is identified by the deep neural network, which has the robot angles as its input. Simulations and experimental results demonstrate that compared with the Back Propagation neural network (BPnn), RBF neural network (RBFnn), Support Vector Machine (SVM) and Extreme Learning Machine (ELM), MLP neural network has the best identification accuracy. And the filters we used can overcome the problem of speed pulse, to some extent, and achieved an ideal imitation result.