OpenNavMap: Structure-Free Topometric Mapping via Large-Scale Collaborative Localization

TL;DR

OpenNavMap introduces a lightweight, structure-free topometric mapping framework for scalable, lifelong visual navigation across multi-session and heterogeneous devices. By combining a three-layer map (Covisibility, Odometry, Traversability) with on-demand 3D geometry from geometric foundation models, it achieves sub-meter localization without pre-built dense maps. The method weds topological localization with DP-based sequence matching, geometric verification, and a confidence-calibrated global optimization, followed by node culling and edge updating to support lifelong operation. Across self-collected and public datasets, including 15.7 km of multi-session data, it achieves robust global consistency (ATE below 3 m) and enables 12 autonomous image-goal navigation trials, demonstrating practical applicability for scalable, crowd-sourced visual navigation with cross-device compatibility.

Abstract

Scalable and maintainable map representations are fundamental to enabling large-scale visual navigation and facilitating the deployment of robots in real-world environments. While collaborative localization across multi-session mapping enhances efficiency, traditional structure-based methods struggle with high maintenance costs and fail in feature-less environments or under significant viewpoint changes typical of crowd-sourced data. To address this, we propose OPENNAVMAP, a lightweight, structure-free topometric system leveraging 3D geometric foundation models for on-demand reconstruction. Our method unifies dynamic programming-based sequence matching, geometric verification, and confidence-calibrated optimization to robust, coarse-to-fine submap alignment without requiring pre-built 3D models. Evaluations on the Map-Free benchmark demonstrate superior accuracy over structure-from-motion and regression baselines, achieving an average translation error of 0.62m. Furthermore, the system maintains global consistency across 15km of multi-session data with an absolute trajectory error below 3m for map merging. Finally, we validate practical utility through 12 successful autonomous image-goal navigation tasks on simulated and physical robots. Code and datasets will be publicly available in https://rpl-cs-ucl.github.io/OpenNavMap_page.

Figures (20)

• Figure 1: Conceptual illustration of the structure-free topometric map generated by our approach, highlighting the traversability layer. During deployment, the robot utilizes RGB-D sensing to traverse the graph derived from multi-session mapping. Green nodes denote the computed shortest path to the goal image, while white nodes represent the global network of traversable locations.
• Figure 2: Block diagram illustrating the pipeline of the proposed OpenNavMap system. The system builds on the topometric map with multiple layers for different utilities: covisibility, traversability, and odometry. a) Individual disconnected submaps are constructed from data collected from various devices. b) The collaborative localization consists of two steps to compute the relative transformation between the reference and query map, and then perform the PGO to jointly estimate their transformation. c) The map merging and updating module fuses these two maps by pruning data with litter information contribution and augment the connectivity of edges to update the map. d) The resulting topometric map is deployed for image-goal navigation.
• Figure 3: Three types of trajectories occurring in multi-session mapping, with difference matrices for the latter two cases in (a). Red dots mark matched loops. The first matrix (Oxford RobotCar maddern20171) shows substantial overlaps, while the second (our dataset; see Sec. \ref{['sec:exp_datasets']}) exhibits sparser, irregular overlaps, highlighting the need for flexible loop closure methods.
• Figure 4: The raw and calibrated confidence maps of the query image are shown after solving the optimization problem \ref{['equ:global_alignment_irls']}. Confidence values are lower in regions not covered by reference images. These values are reweighted based on the residual error, resulting in a greater reduction in low-confidence regions (e.g., $0.44\rightarrow 0.35$) compared to high-confidence regions (e.g., $0.64\rightarrow 0.59$).
• Figure 5: Illustration of strategies for cross-device localization. (a) shows how a panoramic equirectangular image, prevalent in databases such as Google Street View, is converted into a synthetic perspective image. This transformation generates a virtual pinhole view with a user-defined FoV and camera intrinsics. (b) shows that the IQA score can quantitatively evaluate quality of an images.