Abstraction Multimodal Low-Dimensional Representation from High-Dimensional Posture Information and Visual Images

Abstract

Imitative learning is an effective method for robots to obtain a novel movement from a person demonstrating many kinds of movement. Many problems need to be solved, however, before a robot can achieve imitative learning. One problem is how to convert visual information on the demonstrator’s motion to kinematic posture information for the learner. This is referred to as a correspondence problem and we have focused on this problem in this study. To solve it, we focus on the formation of a low-dimensional representation that integrates sensory information from two different modalities. We propose a computation method for constructing the low-dimensional representation combining posture information and visual images by using Kernel Canonical Correlation Analysis (KCCA). Using this method, a robot becomes able to estimate posture information from visual images in a bottom-up way. Using several experiments we show how effective our proposed method is in estimating kinematic information.