A CNN Approach to Assess Environment Complexity for Robotics Autonomous Navigation

Abstract

Growing interest exists in the evaluation of mobile robots performing autonomous navigation. The environment where the robot operates plays an important role in the successful execution of its autonomous mission. It is therefore crucial to assess the complexity of the environment where the vehicle is deployed. In this paper, we identify two parameters which represent meaningful metrics for the evaluation of how challenging a 2D environment is for autonomous navigation. We show how these two parameters can be estimated with a CNN architecture, given as input only a map of the environment. The method is validated on two different datasets and proves successful in achieving very accurate prediction results.