Clustering of motion data from on-body wireless sensor networks for human-imitative walking in bipedal robots

This paper presents an alternative inexpensive and rapid approach for programming behaviour in commercial off-the-shelf bipedal robots. It combines on-body wireless sensor networks to capture human motion and unsupervised learning algorithms to identify key features in human motion. This paper compares three unsupervised learning algorithms for the classification of motion data from an on-body orient motion capture system for training the KHR-1 bipedal robot. The results of the clustering were first compared in the Webots simulator and promising candidates were transferred to the real robot and the results of the experiments have been presented. The EM clustering algorithm worked best and the reason for this have been analysed.