Unsupervised embrace pose recognition method for stuffed-toy robot

This paper presents a new approach to model and recognize embrace interaction based on an embrace-comfortable fabric-based touchpad attached to a stuffed-toy robot with <i>k</i>-mean clustering of location-based features. Evaluation of the method demonstrated its ability to recognize embrace poses in new users with a suitable number of clusters. Consequently, the proposed class assignment method, which assigned classes based on the most common patterns, was used to determine the number of clusters in the model selection experiment. The experimental results showed that the selected model could obtain sufficient recognition performance, which contributed to a guideline for model selection, based directly on variations of embraces without requiring ground truths. This guideline could potentially be applied to different target populations and definitions of the embrace pose.