Rough Terrain Perception Through Geometric Entities for Robot Navigation

Abstract

This paper presents the implementation of a nonlinear geometric cost function to be used with a learning to search algorithm (LEARCH) to robot navigation in rough terrains. The non-linear function introduced is a neural network trained with geometric entities as inputs (points, lines, spheres, planes). These inputs were codified using the Conformal Geometric Algebra framework in order to describe the features of the rough environment where the robot is going to navigate. The geometric entities contain implicitly more information about rough terrain that simple features obtained with image edge-detectors, furthermore by using them as descriptors, the dimension of the feature space is greatly reduced with regard to the dimension of features obtained with sophisticated feature detectors as SIFT or SURF. The advantages of using geometric entities with LEARCH algorithm are shown in the experimental results section of this paper.