Lidar Based Autonomous Navigation of an Unmanned Ground Vehicle Using Statistical and AI Based Steering

In this paper, we study pure lidar based autonomous navigation of an unmanned ground vehicle (UGV) and its experimental verification. The UGV used in this study is a differential wheeled robot, and the main focus is not a SLAM type localization and mapping, instead it is a lidar based surface classification and driveable surface detection. Our test environment is relatively flat and contrary to the recently published results on slope estimation, we propose the use of statistical and/or AI based tools for surface type detection, like sidewalk, vegetation, sand, etc and autonomous steering based on a simplified sensory processing. The proposed statistical approach is computationally lightweight, and performed quite well experimentally. We also proposed an AI based approach for surface classification using synthetic data and obtained promising results. More precisely, steering angle error of around 5 degrees is achieved by using a small AI model of only 22K paramaters. As future work, authors plan to do more experimental tests on surface type classification using AI models.