Point2Radio: AI Foundation Model for Cross-Scene Radio Field Prediction

Researchers have introduced Point2Radio, a foundation model designed to predict radio fields across diverse environments using material-aware point clouds. The model learns a transferable propagation prior from multiple scenes, enabling fast inference on new scenes without the need for meshes or explicit path tracing. Point2Radio employs a common encoder to produce a transmitter-conditioned scene representation, which can be queried at arbitrary receiver locations. Task-specific decoders map this representation to various radio quantities, including three-dimensional path-gain fields and power angular spectra. The model's performance was evaluated on a scene-disjoint split of a corpus containing 337 scenes and 86,272 samples. This approach addresses the inefficiency of simulating radio fields for each scene individually, offering a scalable solution for applications in wireless communications and sensing. The paper is available on arXiv under the identifier 2607.28994.