Probabilistic self-localization for mobile robots

Abstract
Localization is a critical issue in mobile robotics. If a robot does not know where it is, it cannot effectively plan move- ments, locate objects, or reach goals. In this paper, we describe probabilistic self-localization techniques for mobile robots that are based on the principle of maximum-likelihood estimation. The basic method is to compare a map generated at the current robot position with a previously generated map of the environment in order to probabilistically maximize the agreement between the maps. This method is able to operate in both indoor and outdoor environments using either discrete features or an occupancy grid to represent the world map. The map may be generated using any method to detect features in the robot's surroundings, including vision, sonar, and laser range-finder. We perform an efficient global search of the pose space that guarantees that the best position is found according to the probabilistic map agreement measure in a discretized pose space. In addition, subpixel local- ization and uncertainty estimation are performed by fitting the likelihood function with a parameterized surface. We describe the application of these techniques in several experiments, including experimental localization results for the Sojourner Mars rover. I. INTRODUCTION Mobile robots must have some method by which to deter- mine their position with respect to known locations in the en- vironment in order to navigate effectively and achieve goals. This is called the localization problem. The most common and basic method for performing localization is through dead-reck- oning. This technique integrates the velocity history of the robot over time to determine the change in position from the starting location (see, for example, (1) and (2)). Unfortunately, pure dead-reckoning methods are prone to errors that grow without bound over time, so some additional method is necessary to pe- riodically correct the robot position. It is common to combine the additional localization technique, such as triangulation from landmarks or map matching, with dead-reckoning using an ex- tended Kalman filter to probabilistically update the robot posi- tion. In this paper, we describe a technique that performs lo- calization infrequently to update the position of the robot. In order to perform localization, we compare a map generated using the robot's sensors at the current position (the local

This publication has 53 references indexed in Scilit: