Learning Interaction Rules through Compression of Sensori-Motor Causality Space

Abstract

A human partner returns a specific response after a robot acts a specific social cue. We define this as interaction rules. Toward natural communication, we focus on social games played between an infant and caregiver. For the social games, we have to solve two problems: (1) the robot have to find the rules in high-dimensional space with a limited number of exemplars; (2) the interaction should not be divided into learning phase and interaction phase. Previous challenges on learning interaction rules didn’t attacked them both ways. In this paper, we solve them simultaneously. The robot calculates sensori-motor causality space utilizing partial correlation analysis, and it compresses the causality space to find the rules in high-dimension utilizing canonical correlation analysis. The experiment of human-robot interaction showed that a real robot with camera and arms (3 DoF for each) learns gesture interaction rules with human in high-dimensional sensori-motor space without dividing interaction into learning/interaction phases.