Leveraging Machine Learning for Terrain Traversability in Mobile Robotics

Abstract

The problem of traversability of soft terrains is hard to solve due to both the inherent modeling complexity and the related compu- tational cost. In this work a surrogate model is used to describe the behavior of soft soil, thus avoiding explicitly simulating it. We leverage machine learning to train a model on real-world data acquired with the Archimede robotic platform in DLR's Moon-Mars test area in Oberpfaffenhofen, Germany. The model is tested using the Gazebo simulation environment by injecting virtual forces that mimic the effect of drift. Results show that the surrogate model shows promise, but showing also noticeable variability, possibly ascribable to the early stage of the model and training dataset.