Autonomous exploration with online learning of traversable yet visually rigid obstacles

Milos Prágr, Jan Bayer, Jan Faigl

Abstract: This paper concerns online learning of terrain properties combining haptic perception with exteroceptive sensing to reason about forces needed to pass through terrains that visually appear as untraversable obstacles. Terrain learning is studied within the context of autonomous exploration. We propose predicting the traversability of potentially obstructing terrains by active perception to establish a connection between the observed geometric environment model and deliberately sampled forces to pass through the terrain using a haptic sensor that probes the terrain in front of the robot. The developed solution uses a Gaussian Process regressor in online learning and force prediction. The robot is navigated by following the information gain to improve traversability and spatial models. The proposed approach has been experimentally verified in fully autonomous exploration with a multi-legged walking robot. The robot is navigated through visually looking obstacles and explores “hidden” areas while following the expected information gain to explore the terrain properties of the mission area.