3D Perception and Environment Map Generation for Humanoid Robot Navigation

Abstract

A humanoid robot that can go up and down stairs, crawl underneath obstacles or simply walk around requires reliable perceptual capabilities for obtaining accurate and useful information about its surroundings. In this work we present a system for generating threedimensional (3D) environment maps from data taken by stereo vision. At the core is a method for precise segmentation of range data into planar segments based on the algorithm of scan-line grouping extended to cope with the noise dynamics of stereo vision. In off-line experiments we demonstrate that our extensions achieve a more precise segmentation. When compared to a previously developed patchlet method, we obtain a richer segmentation with a higher accuracy while also requiring far less computations. From the obtained segmentation we then build a 3D environment map using occupancy grid and floor height maps. The resulting representation classifies areas into one of six different types while also providing object height information. We apply our perception method for the navigation of the humanoid robot QRIO and present experiments of the robot stepping through narrow space, walking up and down stairs and crawling underneath a table.KEY WORDS-humanoid robot navigation, 3D environment perception, range image segmentation, stereo vision

OctoMap: an efficient probabilistic 3D mapping framework based on octrees

“…Whenever there is a single surface that the robot uses for navigation, an elevation map is sufficient to model the environment, since overhanging obstacles that are higher than the vehicle, such as trees, bridges or underpasses, can be safely ignored. The strict assumption of a single surface can be relaxed by allowing multiple surfaces per cell (Triebel et al, 2006;Pfaff et al, 2007), or by using classes of cells which correspond to different types of structures (Gutmann et al, 2008). A general drawback of most 2.5D maps is that they do not represent the environment in a volumetric way but discretize it in the vertical dimension based on the robot's height.…”

From 3D point clouds to climbing stairs: A comparison of plane segmentation approaches for humanoids

“…Their humanoid autonomously climbed a single step. Gutmann et al [12] use the efficient scan-line grouping algorithm [2] discussed in this paper to extract models of steps given stereo data. Using such models, the humanoid QRIO was able to climb up and down staircases consisting of four steps.…”

“…The approach was originally presented by Jiang and Bunke [18] and performed best in a comparison with other earlier range segmentation techniques [19]. Since then the algorithm has been extended and improved in order to deal with range data containing varying levels of noise, such as range images obtained from stereo vision [2]. The method has successfully been applied on Sony's QRIO robot, allowing the humanoid to recognize Fig.…”

“…These approaches have been successfully applied to humanoid navigation on nonflat ground in the past. The first approach was presented by Gutmann et al [2]. It is based on scan-line grouping and first extracts line segments from neighboring scan lines in a range image.…”

Real-time replanning using 3D environment for humanoid robot

“…Previous studies set several hypotheses to reduce the complexity to guarantee the real-time operation, for example restricting the obstacles to 2D shapes [1] or simple geometries [2]. Recently interactive 3D navigation by humanoid [3] [4] has been reported, but it is for static environments.…”