Segmented SLAM in Three-Dimensional Environments

Abstract

Simultaneous localization and mapping (SLAM) has been shown to be feasible in many small, two-dimensional, structured domains. The next challenge is to develop real-time SLAM methods that enable robots to explore large, three-dimensional, unstructured environments and allow subsequent operation in these environments over long periods of time. To circumvent the scale limitations inherent in SLAM, the world can be divided up into more manageable pieces. SLAM can be formulated on these pieces by using a combination of metric submaps and a topological map of the relationships between submaps. The contribution of this paper is a real-time, three-dimensional SLAM approach that combines an evidence grid-based volumetric submap representation, a robust Rao-Blackwellized particle filter, and a topologically flexible submap segmentation framework and map representation. We present heuristic methods for deciding how to segment the world and for reconstructing large-scale metric maps for the purpose of closing loops. We demonstrate our method on a mobile robotic platform operating in large, three-dimensional environments