A note on robot localization Simon Parsons Department of Computer and Information Science Brooklyn College, City University of New York Brooklyn, NY 11210 parsons@sci.brooklyn.cuny.edu January 10, 2005 Abstract This is a brief note on the localization techniques developed by Wol- fram Burgard, Dieter Fox and colleagues, and used by the MetroBots RoboCup Four-legged league team. In particular, we describe the tech- niques of Markov localization and Monte Carlo localization, summarize the most important publications that describe them, and identify the var- ious flavours of these techniques outlined in those publications. 1 Introduction The aim of this note is twofold. First it is intended as a summary of my understanding of the probabilistic techniques for localization—establishing the location of a robot from sensor data—that we have used in the MetroBots team. As such it is intended as a source of reference descriptions against which our implementations can be checked, and as record of the variations on the basic technical theme that can be found in the literature. Second, it is intended as a guide, albeit a very specific one, to the literature that describes these techniques. As a guide this note makes no pretentions to be comprehensive— it focuses exclusively on the work of Burgard, Fox and colleagues—and leans heavily towards those aspects of the techniques that are most applicable to the RoboCup task. However, it should still be of use to those interested in localizing in other mobile robot scenarios. 2 Basic theory The basic schema for these probabilistic localization techniques is as follows. The description is gleaned from [18]). Figure 1 (borrowed from [16], indeed from the chapter written by Sebasian Thrun) represents the general schema for localization as a dynamic Bayesian network—based on the pose X t-1 of the robot at time t-1, and the action A t-1 it carries out at that time, we can predict 1