Radiance-Field Guided Pretraining: Scaling Localization Models with Unlabeled Wireless Signals

Radio frequency (RF)-based indoor localization offers significant promise for applications such as indoor navigation, augmented reality, and pervasive computing. While deep learning has greatly enhanced localization accuracy and robustness, existing models still struggle with cross-scene generalization because they rely on scene-specific labeled data. To address this, we introduce Radiance-Field Guided Pretraining (RFGP), a novel self-supervised pretraining framework. Unlike conventional pretraining approaches, RFGP utilizes a physics-guided asymmetric autoencoder architecture. Specifically, a large localization model (LM) encodes received RF spectra into latent representations, while a neural radio-frequency radiance field (RF-NeRF) acts as a physics-informed decoder to reconstruct the original spectra via ray tracing. This unique design encourages the LM to learn more transferable representations from unlabeled RF data, which can be collected continuously with minimal effort. To evaluate RFGP, we constructed a massive-scale dataset comprising 7,327,321 RF samples across 100 diverse scenes using four wireless technologies (RFID, BLE, WiFi, and IIoT). Extensive evaluations on unseen test scenes show that the RFGP-pretrained LM reduces localization error by over 40% compared to training from scratch and by 21% compared to standard supervised pretraining, underscoring the potential of RFGP as a scalable pretraining framework for generalized indoor localization. The dataset and implementation code are available in our repository ¹ .