WayTrace: urban navigation from map to motion
Our new paper WayTrace: A Map-to-Language-to-Trace Hierarchy for Embodied Urban Navigation is now available on arXiv.
WayTrace turns turn-by-turn directions from a web map into language commands for a vision-language model, which predicts a trace in image space that a learned navigation policy follows. We deployed it on a wheeled-legged robot over long-range missions through city centres, including crossings, stairs and construction sites. I co-authored the work with Manthan Patel, Jonas Frey, Marco Hutter and Cesar Cadena.