Purely Topological Exploration of Underground Environments

Underground environments pose unique challenges for robotic exploration due to their large scale, sensing degraded nature, and presence of sections with little distinctive features as is the case with smooth tunnels. To address these issues, we propose an exploration technique for underground environments based on a purely topological understanding of the environment, suitable for such featureless scenarios. Instead, a limited set of topological structures (galleries, end of galleries and intersections) are the only distinguishing characteristics used for navigation and mapping. The method is built around a Convolutional Neural Network (CNN) that process the readings of a Velodyne VLP-16 LiDAR, and provides an estimation of where tunnels are distributed around the robot. From these predictions, we build a navigation system that can interpret high-level instructions like “take the next left turn” or “continue straight at the intersection” and on top of this navigation method, we build a mapping system that can explore an underground environment (network of tunnels) in a purely topological fashion. The system is tested in simulation, both in synthetic environments, and a reconstruction of a real tunnel, and proves capable of mapping them under the condition that no loops are present.