Fast online collision avoidance for mobile service robots through potential fields on 3D environment data processed on GPUs

Abstract

This paper demonstrates the fitness of massively parallel exact Euclidean distance transform (EDT) computation for fast 3D online motion planning with potential field and wavefront planners. We combine point-cloud sensor data to gather detailed 3D voxel maps of complex environments. Unlike other approaches, we do not use 3D polygon meshes or reduce the environment to 2D or 2.5D models to improve planning times. The evaluation shows that fast sub-second local and global planning times on the basis of GPU EDT are possible.