AGGC: Anchor-Guided Geodesic Charting for Few-Label CSI Indoor Localization

Abstract

Indoor localization based on channel state information (CSI) exploits waveform-dependent spatial propagation signatures of wireless channels and is an important task for integrated sensing and communication in sixth-generation mobile communication systems. However, existing CSI fingerprint localization methods usually require dense coordinate labels, which limits low-cost deployment. While channel charting exploits unlabeled CSI to learn relative spatial structures, coordinate anchors are typically used only for post-hoc affine alignment. To address this limitation, we propose Anchor-Guided Geodesic Charting (AGGC), which incorporates sparse anchors into both geodesicprior calibration and coordinate-supervised learning. This design improves the reliability of the learned spatial structure and enables direct coordinate prediction. Experiments on the opensource DICHASUS indoor dataset show that AGGC achieves stable localization performance under different label budgets and environments. Specifically, on the cf0x dataset with 10% coordinate anchors, AGGC reduces the mean localization error from 0.544 m to 0.299 m and the 90th-percentile error from 0.968 m to 0.534 m over the ADP-Geodesic baseline, demonstrating its effectiveness in few-label CSI indoor localization. The main experimental code will be made publicly available at github.com/cjr22/AGGC/.