Prediction of Terrain Traversability Cost Using Self-supervised Learning for Mobile Robot in Uneven Terrain

Abstract

In this paper, we propose a method to predict terrain traversability cost using self-supervised learning for mobile robot in uneven terrain. The robot learns to predict risk of traversing the terrain by using the vibration information from the acceleration sensor as the prediction target and the color and depth image from the camera as the input. We conducted an experiment to get the datasets in a real environment and predict terrain traversability cost from terrain image at constant speed. The prediction results showed that low and medium cost could be predicted. In addition, we propose safer prediction method considering speed change.