Point2Radio: A Foundation Model for Cross-Scene Radio Fields from Material-Aware Point Clouds

📄 Paper (arXiv): Point2Radio
🌐 Interactive demo: [TBD]

---

What it does

Given a material-aware point cloud and a transmitter (TX) location, Point2Radio learns a transferable scene representation that can be queried at arbitrary receiver (RX) positions to predict:

• 3D path-gain (PG) fields — dense spatial coverage in a new room, in one feedforward pass
• Power angular spectra (PAS) — directional arrival spectra via a task-specific decoder on the same encoder

At inference it needs only the point cloud + transceiver queries (milliseconds on a single GPU). No meshes and no online ray tracing. The same backbone can be lightly adapted to a target scene when a few labels are available.

Path Gain (PG)

Power Angular Spectrum (PAS)

---

Code & data

Full code, pretrained checkpoints, and dataset release are coming soon.

---

Citation

@misc{wen2026point2radiofoundationmodelcrossscene,
      title={Point2Radio: A Foundation Model for Cross-Scene Radio Fields from Material-Aware Point Clouds}, 
      author={Chaozheng Wen and Chenghong Bian and Hongze Chen and Jun Zhang},
      year={2026},
      eprint={2607.28994},
      archivePrefix={arXiv},
      primaryClass={cs.NI},
      url={https://arxiv.org/abs/2607.28994}, 
}