CVGL-SLAM: Satellite Imagery as an Absolute Constraint for Drift-Bounded SLAM

In preparation 2026

Abstract

Autonomous robots increasingly operate across large-scale outdoor environments, where reliable navigation depends on accurate self-localization. Odometry provides this estimate but drifts over distance. Existing approaches to bound this drift are limited: loop closure needs revisits, GNSS fails in canyons, and prior maps are costly to build. We introduce CVGL-SLAM, which uses cross-view geo-localization (CVGL), matching ground observations to globally available satellite imagery, as an absolute constraint on odometry. To use noisy cross-view matches, it keeps only confident ones, and anchors the unconstrained vertical axis with gravity. CVGL-SLAM bounds odometry drift at large scale and outperforms odometry-only and loop-closure baselines.

Factor graph. LIVO odometry links consecutive poses. Confident CVGL matches tie each pose to the anchor Ta (world to UTM); matches with confidence below 0.3 are rejected.
Factor graph. LIVO odometry links consecutive poses. Confident CVGL matches tie each pose to the anchor Ta (world to UTM); matches with confidence below 0.3 are rejected.

Results on TartanDrive 2.0

Table 1. Absolute trajectory error (3D ATE RMSE, m) on TartanDrive 2.0, aligned on the first 100 m. Methods marked † use loop closure; ours does not. Best in bold.

Method1253131113421809pt2
4.4 km2.2 km1.7 km1.6 km1.6 km
LTA-OM† 1.91 324 2.33 1.44 9.13
GLIM† 3.00 17.3 5.71 fail 2.99
Voxel-SLAM† 3.86 8.92 3.25 2.78 4.19
SE(3)-LIO 2.74 14.3 2.99 1.40 6.37
Ours 2.60 1.52 1.59 1.39 1.86
TartanDrive 2.0 sequences on satellite imagery, 12x speed. Blue: CVGL-SLAM. Orange: SE(3)-LIO odometry. Dots: CVGL matches, colored by confidence.