Evidence Meshes for Three-Dimensional Modeling, Visualization, and Navigation

As robots are tasked with ever more complex missions, they demand more sophisticated models of the environments in which they must work. Rough-terrain mobility, site surveying, and dexterous manipulation all demand a fully 3D map of the world that simultaneously exhibits large scale and high resolution, a situation we refer to as scale disparity. Most robots discretize the world into a uniform-grid that is used to accumulate evidence from multiple measurements. Unfortunately the memory footprint of such maps grows dramatically with scale disparity. Octrees can lessen memory requirements, but do not fully counteract the exponential growth of the underlying grid representation. In response, we are developing a map representation called an Evidence Mesh that provides the benefits of probabilistic treatment of evidence but performs better under scale disparity. It is based on a triangulated mesh and is compatible with well-known simplification algorithms to represent the shape of objects at adjustable levels of fidelity. Like an evidence grid but unlike other mesh-based mapping methods available today, an Evidence Mesh accumulates evidence about the location of objects through simplification and across multiple sensor measurements, enabling robust noise filtering and avoiding artifacts and aliasing introduced by artificial grid structures.