Graph Neural Networks for RFID-Based Spatial Geometry Inference in Spatial AI Systems

This paper addresses the limitations of traditional RFID localization, which typically treats sensor readings as independent, noisy data points. Instead, the authors propose a relational framework that models RFID observations as nodes in a graph. By incorporating floorplan semantics—such as room boundaries, antenna locations, and storage containers—the model transforms raw signal strength (RSSI) and coordinate data into a structured representation. This allows a Graph Neural Network (GNN) to perform spatial reasoning, shifting the goal from simple coordinate estimation to the inference of geometric patterns like movement trajectories and bounded occupancy regions.

In many indoor environments, RFID infrastructure is deployed for inventory management rather than dedicated geometric sensing. This research shows that such infrastructure can be repurposed as a powerful spatial sensing field. By leveraging the relational structure of the environment, the proposed GNN approach provides a more robust way to interpret sparse and noisy data. This is particularly valuable for Spatial AI systems, such as digital twins or smart building automation, where understanding the spatial context of object movement is as important as knowing an object's precise location.

The authors developed a pipeline that parses floorplan files to enrich RFID detections with semantic context. Observations are grouped into temporal windows, and features are engineered to include distances to infrastructure and semantic zones. The model was evaluated on data from a controlled laboratory environment, demonstrating that graph-based learning can successfully categorize spatial patterns. The results suggest that topological consistency, derived from the environment's physical layout, acts as a strong prior that helps the model overcome the inherent ambiguity of individual RFID signal measurements.