Data-driven NAO Robot Joint Movement Control

Imitating and learning human motion, and eventually realizing independentbehavior is an important research subject of robot intelligence. Based on NAO robotplatform, this thesis studies the data-driven methods of robot joint movement. Themain work includes the following two aspects.According to one-to-one mapping of human-robot biological model, this thesisdesigns the experiment schemes that body left shoulder data drive NAO robotshoulder joint. Based on MTi sensors worn on the left arm, human joint motioninformation is accessed in real-time, serial port communication is used to send data toPCs. The application codes extracting and calculating data is programmed on VisualStudio2010platform. The data is converted to Roll, Pitch and Yaw forms that NAOrobots can understand. Through a wireless network, the connection is set up betweenPC and NAO. Based on Choregraphe, NAO is driven by motion data to imitate thehuman’s joint actions. Experiments show that NAO can thoroughly imitate humanbody joints route.Based on echo state network, the central pattern generator (CPG) model isdesigned and trained in MATLAB. With MTi sensor, human right arm and left legmotion data in walking are collected to train echo state network and build the CPG.Taking the leg movement data as the teacher signal, the CPG model is trained forcompletely autonomous motion. Experiment results show that this method has someflaws in phase tracking. Then Euler and acceleration data of arms are used as inputrespectively, leg movement data as output, to train the CPG model. By contrast, theEuler data has better tracking performance on amplitude, acceleration data better onphase.This thesis is a primary research. The problems such as debugging NAO robot,and the movement data modeling based on echo state network, need further researchin the future.