Learning by demonstration from multiple agents in humanoid robots

摘要

In this paper we present a novel hand gesture recognition system for adaptive learning of human interactions for intuitive human-robot interactions. The proposed system tracks upper body hand gestures of two interacting human agents simultaneously with the help of a motion tracking camera. High dimensional spatio-temporal gesture data is modelled onto low dimensional latent space by means of principal component analysis (PCA) yet maintaining maximum spatial information. Mapping between these low dimensional motion data is learned by non-linear learning algorithms like Radial Basis Function Neural Network (RBFN) and Long Short Term Memory Neural Network (LSTM). Finally learned interaction models can be used by robot to perform similar interactions with a human partner. In order to evaluate accuracy of gestures performed by the robot, the paper presents novel approach to find optimum number of neurons required in the hidden layer of a three layered artificial neural network to produce the desired gesture. Gestures are performed by both, a simulated virtual robot and a physical humanoid robot platform.