A study of learning and automatic motion generation with emotional factors

Abstract

The research focuses on a technique to learn the basic motions and an emergence of new motions with emotions for entertainment robots. The proposed system consists of hierarchical sandglass-type neural networks (SNNs) and emotional factors attached network (EFAN) which transforms learned motions into emotional motions. A SNN can work as a non-linear PCA (principal component analysis). Each SNN extracts position features from joint angles of a robot and motion features from time-series position features. The trained SNNs can reconstruct the learned motions corresponding to the motion features. Emotions are assumed to be expressed by some emotional factors such as speed, boldness, facial expression and so on. As framing EFAN using the parameters of emotional factors and motion features, EFAN generates emotional additions applied to basic motions. The proposed system generates basic motions and emotional motions with feature parameters.