Simulated dataset released with the NeurIPS 2026 paper "RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis" by Kang Yang and Mani Srivastava.

📄 Paper: arXiv:2605.24290 · 💻 Code: github.com/nesl/RxGS · 🤖 Models: kyang73/RxGS-pretrained

It pairs 5,089 transmitter (TX) positions with 21 receivers (RX) in one scene, for training and evaluating RF data synthesis at any transmitter and any receiver.

spectrum_multirx.tar holds 106,869 spatial spectra (5,089 TX × 21 RX), simulated with Sionna RT.

The scene is a cluttered 8 m × 6 m × 3 m conference room with ITU-material walls and furniture, at 2.4 GHz. For each TX–RX pair, Sionna ray tracing (up to 5 bounces) computes the channel at an 8 × 8 receiver array, and conventional beamforming over a 1° azimuth–elevation grid gives a peak-normalized 90 × 360 grayscale spectrum image. Receivers are at random positions at z = 2 m. Files: spectrum/<tx>_<rx>.png, tx_pos.csv, rx_positions.yml. The code splits the TX positions 80/20 at random (seed 8371): 4,071 train and 1,018 test.

The paper's other two datasets are not hosted here: BLE RSSI comes from NeRF², and WiFi CSI is built from the KU Leuven Ultra-Dense Indoor MaMIMO CSI Dataset with the preprocessing script in the code repository.

Download

From the root of the code repository:

hf download kyang73/RxGS-data spectrum_multirx.tar --repo-type dataset --local-dir data
tar -xf data/spectrum_multirx.tar -C data && rm data/spectrum_multirx.tar

This gives data/spectrum_multirx/. The images ship as one archive because the folder holds more files than the Hub allows per directory.

Citation

@inproceedings{yang2026rxgs,
  title     = {RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis},
  author    = {Yang, Kang and Srivastava, Mani},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}