Category-Based Intrinsic Motivation

Abstract

A goal of epigenetic robotics is to design a control architecture that implements an ongoing, autonomous developmental process which is unsupervised, unscheduled, and taskindependent. The developmental process we are currently exploring contains three essential mechanisms: categorization, prediction, and intrinsic motivation. In this paper we describe a hybrid approach that uses Growing Neural Gas for categorization, neural networks for prediction, and Intelligent Adaptive Curiosity for intrinsic motivation. We apply this system to a physical robot operating in a dynamic visual environment and analyze the types of categories it forms.