RangeNav: LiDAR range-image pretraining for navigation on arXiv
Our paper RangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation is now available on arXiv.
RangeNav is a compact (183k parameters) LiDAR range-image encoder pretrained on synthetic data with navigation-aligned objectives, then frozen for policy learning. It was deployed on a wheeled-legged robot in dynamic crowds, low-light and construction environments. Joint work with Samuel Foo Enze, Manthan Patel, Fan Yang, Linus Kramer, Marco Hutter and Cesar Cadena.