Evaluation of SLAM methods and adaptive Monte Carlo localization

Abstract

Mobile robots that perform tasks autonomously must be able to solve two fundamental problems on their own to ensure autonomy. On the one hand, a map of an unknown environment must be created and the robot’s position determined simultaneously, for which the Simultaneous Localization and Mapping (SLAM) solution is used. On the other hand, an autonomous localization algorithm, such as Adaptive Monte Carlo Localization (AMCL), must be used to solve the kidnapped-robot problem. Every sensor has some measurement inaccuracy, which in the case of a laser distance sensor (LDS) is very small and has no significant effect on the accuracy of the generated map, but this inaccuracy adds up with each measurement until there is an offset in the map and makes it unusable. Therefore, SLAM algorithms must be able to produce a detailed map and cover a wide range of LDS with various measurement inaccuracies as well. The localization algorithm AMCL, in contrast, shows major limitations in localization after the occurrence of the kidnapped-robot problem, if the position of the robot has been changed significantly, or several particles have already formed a swarm and represent a possible position of the robot. For this reason, AMCL is often used only as a base for improved localization algorithms. In this work, we evaluated how robustly and reliably the three SLAM methods Gmapping, Hector, and Karto perform when using LDS with different technical characteristics. In addition, an optimized parameterization of AMCL was performed and adapted to the environment to ensure autonomous localization after the occurrence of the kidnapped-robot problem, without the need for external intervention.This work was carried out in a simulation in the Robot Operation System (ROS), where an apartment was created as a simulated world. Maps of this virtual environment were created with different technical characteristics of the LDS, which were compared to the floor plan of the apartment using a program developed for this work. This allowed us to determine how the different inaccuracies of the LDS affected the accuracy of the map. For the evaluation of AMCL with the adjusted parameters, different scenarios simulating the Kidnapped-Robot problem with different difficulty levels were performed. In order not to reduce the efficiency of this localization in the presence of small position shifts, the referencing process was divided into three phases, which allowed for fast localization even in the presence of small position changes.The results of this work show that both Gmapping and Karto are very robust and reliable SLAM methods that produce usable maps even in the presence of high inaccuracies in the LDS, while Hector has problems determining the position of the robot even in the presence of small deviations, resulting in an overlap in the map. Furthermore, it has been shown using various kidnapped-robot scenarios that AMCL is always able to perform localization, provided that the parameters are adapted to the environment.