OpenNavMap: Multi-Session Structure-Free Topometric Mapping for Scalable Visual Navigation

Jianhao Jiao, Changkun Liu, Jingwen Yu, Dimitrios Kanoulas

Keywords: Autonomous Navigation, Visual Localization, Collaborative Localization

TL;DR: We present OPENNAVMAP, a collaborative multi-session mapping system for visual relocalization (VLoc) and navi-gation (VNav).

Abstract: This paper proposes OpenNavMap, a multi-session mapping system designed for scalable visual navigation. Rather than relying on the 3D structure-based representation of the environment, OpenNavMap adopts a robust collaborative lo-calization strategy to facilitate map merging, taking only 2D images as input. The resulting topometric map is thus lightweight and structure-free, composed of three layered graphs: odometry, covisibility, and traversability. This design enables autonomous visual navigation without the need for prior structure-based maps. Experiments on map merging demonstrate that OpenNavMap achieves high accuracy (< 3m ATE over 15km) and strong robustness to challenging conditions such as day-night transitions and large viewpoint changes. The system has been successfully deployed on a quadruped robot using only monocular RGB inputs for image-goal visual navigation. A video (https://drive.google.com/file/d/1bFKZstoTOoO_OOAB6hvKeVK5O_q2e-zq/view?usp=drive_link) is provided to explain the methodology and experimental results.

Submission Number: 4