Self Map: Self-Supervised Radio Mapping for Zero-Calibration Indoor Positioning
Abstract
Indoor positioning systems have achieved meter-level accuracy; however, their deployment is still prohibitive due to the high cost of manual calibration in site surveys, fingerprint collection, and infrastructure annotation. We present SelfMap, a self-supervised learning framework that allows a mobile device to localize itself in an unseen building without any labeled data from the target environment. SelfMap learns a latent radio map by leveraging the natural geometric consistency between physical motion measured by on-device Inertial Measurement Unit (IMU) and signal space transitions from ambient Wi-Fi and Bluetooth. The insight is that physical displacement provides free supervision: fingerprints from nearby locations are similar, and this relationship can be learned without manual labeling. SelfMap applies three complementary self-supervised losses - temporal consistency, geometric preservation, and signal reconstruction - to learn a robust device-invariant embedding. We evaluate SelfMap in three different types of buildings (office, university, and shopping mall) and demonstrate that SelfMap achieves a mean localization error of 2.4 meters with zero calibration points versus 1.8 meters for a fully supervised baseline requiring several hours of labeled data collection. SelfMap also generalizes across buildings with less than 0.5 m accuracy degradation. Our results indicate that self-supervised learning may indeed remove the main barrier to the large-scale deployment of IPS.