[Paper] PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting
Source: arXiv - 2605.08035v1
Overview
PropSplat introduces a map‑free way to reconstruct radio‑frequency (RF) propagation fields using only sparse transmitter‑receiver measurements. By representing the environment with learnable 3‑D anisotropic Gaussian “blobs,” the method sidesteps costly ray‑tracing or exhaustive site surveys, delivering accurate path‑loss predictions for both outdoor and indoor scenarios.
Key Contributions
- Map‑free propagation modeling: No need for floor plans, terrain databases, or clutter maps.
- 3‑D anisotropic Gaussian primitives: Each Gaussian encodes a scalar offset to a baseline path‑loss model, allowing fine‑grained environmental adaptation.
- End‑to‑end learning pipeline: Gaussians are automatically seeded along measured links and jointly optimized with a learnable path‑loss exponent.
- State‑of‑the‑art accuracy: Beats recent wireless radiance‑field approaches (NeRF², GSRF, WRF‑GS+) on large outdoor drive‑test data and indoor BLE localization.
- Sparse data friendliness: Works with measurements spaced as far as 300 m apart, dramatically reducing the measurement burden.
Methodology
-
Baseline path‑loss model – Start with a classic log‑distance model
$$
PL(d) = PL_0 + 10\alpha \log_{10}(d)
$$where the exponent (\alpha) is learnable.
-
Gaussian “splatting” – Populate the 3‑D space with anisotropic Gaussian primitives. Each primitive holds a single scalar value that adjusts the baseline loss along any ray that passes through it.
-
Initialization – For every observed transmitter‑receiver pair, place a Gaussian at the midpoint of the line‑of‑sight (or the measured path). This gives the network a sensible starting point without any external map.
-
End‑to‑end optimization – Using the measured received signal strength (RSS) values, back‑propagate the error through the Gaussian field and the baseline exponent. Stochastic gradient descent tunes both the Gaussian amplitudes and the exponent simultaneously.
-
Inference – To predict RSS at any arbitrary point, the model queries the accumulated contribution of all overlapping Gaussians along the straight line to the transmitter, adds the baseline loss, and outputs the estimated field value.
The whole pipeline runs on commodity GPUs and converges in minutes for datasets of a few thousand measurements.
Results & Findings
| Dataset | Metric | PropSplat | Best Prior (WRF‑GS+) | Gap to Ground Truth |
|---|---|---|---|---|
| Outdoor drive‑test (6 sub‑6 GHz, 300 m spacing) | RMSE (dB) | 5.38 | 5.87 | ~0.5 dB improvement |
| Outdoor – same data | RMSE vs. NeRF² | 5.38 vs. 14.76 | – | >9 dB gain |
| Indoor BLE (localization) | Mean error (m) | 0.19 | – (NeRF²: 1.84) | ~10× better |
| Indoor BLE (RSS prediction) | RMSE (dB) | ≈0.9 (≈NeRF²) | – | Comparable while being far more accurate for position |
Key Takeaways
- Sparse measurements are sufficient – Even with 300 m gaps, the model captures large‑scale terrain effects and fine‑scale indoor multipath.
- Gaussian primitives efficiently encode clutter – The anisotropic shapes adapt to walls, foliage, or building geometry without ever seeing a map.
- Training is stable – The baseline path‑loss term prevents the optimizer from diverging, while the Gaussians fine‑tune local deviations.
Practical Implications
| Who | Why It Matters |
|---|---|
| Network planners | Deploy a new small‑cell or Wi‑Fi network in a city block without waiting for up‑to‑date GIS data; a quick drive‑test yields a high‑fidelity propagation map. |
| IoT developers | Estimate coverage for low‑power protocols (BLE, LoRa) in warehouses or factories using only a handful of RSS samples, reducing site‑survey costs. |
| AR/VR location services | Leverage the indoor BLE localization accuracy (sub‑0.2 m) for indoor positioning systems that need to work out‑of‑the‑box. |
| Edge‑AI & autonomous vehicles | Integrate PropSplat into on‑board radios to predict dead zones ahead of time, improving handover decisions without a pre‑loaded map. |
| Research & open‑source tools | The method is GPU‑friendly and can be wrapped into existing RF‑simulation libraries, enabling rapid prototyping of new propagation‑aware algorithms. |
In short, PropSplat lowers the barrier to site‑specific RF modeling, turning a traditionally expensive, map‑dependent workflow into a data‑driven, on‑the‑fly capability.
Limitations & Future Work
- Line‑of‑sight assumption for initialization – Gaussians are seeded along straight paths; heavily obstructed environments may need more sophisticated seeding strategies.
- Scalability to city‑scale – While the method handles multi‑region outdoor tests, memory consumption grows with the number of Gaussians; hierarchical or sparse voxel representations could help.
- Frequency generalization – The current formulation learns a single path‑loss exponent per frequency band; extending to broadband or multi‑band joint training is an open question.
- Dynamic environments – Moving objects (vehicles, people) are not modeled; incorporating temporal updates or online learning would make the system robust to real‑time changes.
The authors suggest exploring adaptive Gaussian placement, temporal extensions, and integration with existing GIS layers (when available) to further boost accuracy and efficiency.
Authors
- William Bjorndahl
- Maninder Pal Singh
- Farhad Nouri
- Joseph Camp
Paper Information
- arXiv ID: 2605.08035v1
- Categories: eess.SP, cs.LG
- Published: May 8, 2026
- PDF: Download PDF