Bayesian Learning Experiments with a Khepera Robot

Abstract

. This paper presents a new robotic programming environment based on the probability calculus. We show how reactive behaviours, like obstacle avoidance, contour following, or even light following, can be programmed and learned by the Khepera with our system. We further demonstrate that behaviours can be combined either by programmation or learning. A homing behaviour is thus obtained by combining obstacle avoidance and light following. 1 Introduction We propose a new robotic programming environment, which was tested on a Khepera robot. This system is based on the probability calculus. The choice of probabilities as a formal system allows an easy and rigorous translation of intuitive knowledge into a program. An example is the expression of dependence or independence between variables. In order to program a behaviour, the programmer will first have to state such a priori knowledge about the task at hand. This "seed" of program can then be tuned by confronting it to experimental data wh...