3D Polygonal Mapping for Humanoid Robot Navigation

The traditional environment representation for footstep planning and collision detection for humanoid robots is the 2.5D height map. Although easy to compute and relatively space efficient, height maps have limitations that prevent a humanoid from achieving its full navigational potential in complex real-world environments, e.g., due to lack of explicit representation of walkable surfaces. In this paper, we propose to represent planar surfaces with slanted polygons embedded into 3D Cartesian space, which significantly reduce the memory footprint required for a map. These 3D polygons have explicit boundaries and serve as planable regions for a humanoid for support placement, e.g., footstep planning, and for object placement and manipulation. At the same time, we use the aforementioned 3D polygons for localization within the map while it is being built from sensor data in real time. We hereby combine visual odometry techniques based on plane and edge registration with well-defined polygonal set operations to build an accurate and compact representation of indoor environments with a global ground plane. The result is a geometric map which not only provides an explicit bounded planar surface representation but also allows analytically computed collision and placement information. As our experimental results obtained with the Nao humanoid robot show, we are able to obtain 3D polygonal maps built over extended sequences of RGB-D frames while maintaining an efficient per frame run-time at an especially small memory consumption.