Fast Traversability Estimation for Wild Visual Navigation

Jonas Frey, Matias Mattamala, Nived Chebrolu, Cesar Cadena, Maurice Fallon, Marco Hutter

Keywords: Navigation, Pre-training, Robot Learning, Traversability, Online Learning

TL;DR: Wild Visual Navigation (WVN) leverages pre-trained high-dimensional visual features for self-supervised online traversability learning for robot navigation.

Abstract: This work demonstrates real-world learning and adaptation in the context of traversability estimation. We present a system, Wild Visual Navigation (WVN), which relies on pre-trained high-dimensional features from a self-supervised visual transformer with an online supervision scheme, to achieve on-the-fly traversability learning from a few samples collected in the field. We validate our system with offline experiments and real-world navigation deployments, showing that pre-trained features are fundamental to achieve fast and robust adaptation to new environments.