Common-patterns based mapping for robot navigation

Abstract

Mobile Robot Navigation problem has been extensively studied for decades, but a general solution which suits to various environments remains a challenging topic. One of the popular methods is to build the map and then navigate based on such map. Although most of the map-building approaches, either metric or topological, can efficiently create the map in an unknown environment, they rely on coordinates so that the error in self-pose estimation is unavoidable. In this paper, we alternatively propose a new map-building approach which is especially suitable to mobile robot navigation and does not rely on coordinates. Two key ingredients of the proposed method are (i) the self-organized common-pattern and (ii) the reasoning technique. First the common-patterns are generated in an unsupervised manner by the Self-Organizing and Incremental Neural Networks (SOINN). These patterns are used to incrementally represent the map of environments. The map generated in this manner is called Common-Patterns Based Map (CPM). The CPM is incrementally generated while the robot wandering in the environment. The reasoning technique is proposed to optimize the CPM. The evaluation of the proposed method is done by the experiment of 3D-physical robot simulators (Webots). All environments are the maze. The results show that the CPM is suitable to the navigation with an impressive rate of memory consumption. The loop can be closed successfully. The navigating performance is superior to that of reinforcement learning as it always requires only two episodes.