Advances in positioning and localization technologies, together with the growing capabilities of mobile devices, have enabled applications such as Augmented Reality (AR) and Autonomous Mobile Robots (AMR). Yet current spatial computing solutions still fail under challenging conditions. Wireless positioning systems such as the Global Navigation Satellite System (GNSS) and Ultra-Wideband (UWB) degrade severely in urban areas and complex indoor spaces, where multipath and Non-Line-of-Sight (NLOS) effects introduce ranging errors that undermine accuracy. Self-localization systems based on cameras, LiDAR, and Inertial Measurement Units (IMUs), though free of dedicated infrastructure, are in turn vulnerable to sparse visual features, poor lighting, and rapid or dynamic motion. Achieving seamless and accurate spatial awareness therefore requires either careful handling of the error sources within individual positioning systems or the integration of multiple modalities for mutual compensation. This project addresses both. (1) Foundation Models for Wireless Positioning Systems: Multipath and NLOS effects are the primary cause of GNSS positioning degradation in urban outdoor environments. Although the resulting ranging errors are difficult to model analytically, their patterns are tightly coupled to geographic sites and can therefore be learned from data. The Principal Investigator (PI) has proposed supervised and weakly-supervised frameworks for ranging-error regression using labeled and unlabeled data, respectively [1, 2, 3]. Building on these, the PI will scale the pipelines spatially and temporally by pre-training on large-scale datasets. The target is to create the city-scale city-specific foundation models, like GNSS Pre-trained Transformer for Stockholm (GnssPT-Stockholm). Similar ideas will be investigated for UWB-based indoor localization. (2) Context-aware Spatial Computing: Rather than repeatedly correcting a single modality, it is often more effective to integrate several. When a user enters a transitional area with weak or degraded wireless signals, complementary sensors such as cameras and IMUs can take over the positioning task; once high-quality radio signals become available again in open areas, the system can revert to wireless solutions, which are generally more lightweight than onboard self-localization. This calls for an intelligent context-classification module within the positioning pipeline, allowing each modality to adapt to diverse and dynamic environments. The PI has developed end-to-end frameworks that learn the relative contributions of different positioning measurements directly from a location-related training loss [2, 3], and will build on these to investigate seamless sensor fusion across large-scale, heterogeneous contexts, like Context-Aware Moving Horizon Location Estimation. [1] Weng, X., Ling, K. V., & Liu, H. (2024). PrNet: A neural network for correcting pseudoranges to improve positioning with Android raw GNSS measurements. IEEE Internet of Things Journal. [2] Weng, X., Ling, K. V., Liu, H., Wang, B., & Cao, K. (2026). NeRC: Neural ranging correction through differentiable moving horizon location estimation. ACM/IEEE SenSys 2026. [3] Weng, X., Ling, K. V., Liu, H., & Cao, K. (2024). Towards end-to-end GPS localization with neural pseudorange correction. IEEE FUSION 2024.