Autonomous Robots 12, 287–300, 2002 c  2002 Kluwer Academic Publishers. Manufactured in The Netherlands. Fast, On-Line Learning of Globally Consistent Maps TOM DUCKETT Department of Technology, University of ¨ Orebro S-70182 ¨ Orebro, Sweden tom.duckett@tech.oru.se STEPHEN MARSLAND AND JONATHAN SHAPIRO Department of Computer Science, University of Manchester, Oxford Road, Manchester M13 9PL, UK smarsland@cs.man.ac.uk jls@cs.man.ac.uk Abstract. To navigate in unknown environments, mobile robots require the ability to build their own maps. A major problem for robot map building is that odometry-based dead reckoning cannot be used to assign accurate global position information to a map because of cumulative drift errors. This paper introduces a fast, on-line algorithm for learning geometrically consistent maps using only local metric information. The algorithm works by using a relaxation technique to minimize an energy function over many small steps. The approach differs from previous work in that it is computationally cheap, easy to implement and is proven to converge to a globally optimal solution. Experiments are presented in which large, complex environments were successfully mapped by a real robot. Keywords: simultaneous localization and mapping, concurrent map-building and self-localization, relaxation algorithm, Gibbs sampling, learning and adaptation 1. Introduction Maps are very useful for mobile robot navigation in complex environments, being needed for self- localization and path planning, as well as enabling hu- man operators to see where the robot has been. While successful navigating robots have been developed us- ing pre-installed maps, to operate in unknown envi- ronments a robot needs the ability to build its own maps. However, the sensor information available to the robot is noisy and can produce errors when inte- grated into the map. In particular, the robot’s odom- etry is subject to drift errors caused by factors such as wheel slippage, which can lead to an inconsistent mapping of the environment. To maintain a coherent representation of the environment that can be recon- ciled with future sensory perceptions, some means of maintaining geometric consistency in the map is required. This paper introduces a fast, on-line algorithm for obtaining globally consistent maps. The approach dif- fers from previous work in that it is computationally cheap, easy to implement and is guaranteed to find a solution that is statistically optimal. Our algorithm as- sumes three sources of perceptual information: (i) a place recognition system, (ii) a global orientation obtained from a compass, and (iii) local distance information from odometry. Experiments are presented in which large, complex en- vironments were successfully mapped by a real robot using ultrasonic range-finder sensors for (i), a flux-gate compass for (ii), and uncorrected odometer sensors for (iii). The map representation consists of a topologically connected network of places, where each link is labeled with noisy metric information describing the relative distance and absolute angle between the two places it