Robot training using system identification

Abstract

This paper focuses on developing a formal, theory-based design methodology to generate transparent robot control programs using mathematical functions. The research finds its theoretical roots in robot training and system identification techniques such as ARMAX (Auto-Regressive Moving Average models with eXogenous inputs) and NARMAX (Non-linear ARMAX). These techniques produce linear and non-linear polynomial functions that model the relationship between a robot’s sensor perception and motor response. The main benefits of the proposed design methodology, compared to the traditional robot programming techniques are: (i) It is a fast and efficient way of generating robot control code, (ii) The generated robot control programs are transparent mathematical functions that can be used to form hypotheses and theoretical analyses of robot behaviour, and (iii) It requires very little explicit knowledge of robot programming, therefore end-users/programmers who do not have any specialized robot programming skills can nevertheless generate task-achieving sensor–motor couplings. The nature of this research is concerned with obtaining sensor–motor couplings, be it through human demonstration via the robot, direct human demonstration, or other means. The viability of our methodology has been demonstrated by teaching various mobile robots different sensor–motor tasks such as wall following, corridor passing, door traversal and route learning.