Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generalizable Radio-Frequency Radiance Fields for Spatial Spectrum Synthesis

Kang Yang, Yuning Chen, Wan Du

TL;DR

The paper tackles RF spatial spectrum synthesis across unseen environments by introducing GRaF, which generalizes beyond scene-specific NeRF models through RF Spatial Spectrum Interpolation Theory. It combines a geometry-aware Transformer-based latent RF radiance field with a neural ray tracing algorithm to synthesize spectra conditioned on nearby transmitter spectra. Key contributions include a formal interpolation bound $\mathbf{SS}(\mathbf{P}_t) \approx \sum_i w_i \mathbf{SS}(\mathbf{P}_i)$ with $\epsilon \le K \max_i \|\mathbf{P}_t-\mathbf{P}_i\|^2$, a latent representation $\mathbf{Z}=\mathcal{T}_\Psi(\mathcal{N}_L,\mathbf{P})$, and a neural ray tracer that jointly models amplitude and phase along RF paths. Empirically, GRaF achieves state-of-the-art generalization to unseen scenes/layouts on simulated and real data, and improves downstream tasks such as AoA localization, while avoiding per-scene retraining required by prior NeRF-based RF methods.

Abstract

We present GRaF, Generalizable Radio-Frequency (RF) Radiance Fields, a framework that models RF signal propagation to synthesize spatial spectra at arbitrary transmitter or receiver locations, where each spectrum measures signal power across all surrounding directions at the receiver. Unlike state-of-the-art methods that adapt vanilla Neural Radiance Fields (NeRF) to the RF domain with scene-specific training, GRaF generalizes across scenes to synthesize spectra. To enable this, we prove an interpolation theory in the RF domain: the spatial spectrum from a transmitter can be approximated using spectra from geographically proximate transmitters. Building on this theory, GRaF comprises two components: (i) a geometry-aware Transformer encoder that captures spatial correlations from neighboring transmitters to learn a scene-independent latent RF radiance field, and (ii) a neural ray tracing algorithm that estimates spectrum reception at the receiver. Experimental results demonstrate that GRaF outperforms existing methods on single-scene benchmarks and achieves state-of-the-art performance on unseen scene layouts.

Generalizable Radio-Frequency Radiance Fields for Spatial Spectrum Synthesis

TL;DR

The paper tackles RF spatial spectrum synthesis across unseen environments by introducing GRaF, which generalizes beyond scene-specific NeRF models through RF Spatial Spectrum Interpolation Theory. It combines a geometry-aware Transformer-based latent RF radiance field with a neural ray tracing algorithm to synthesize spectra conditioned on nearby transmitter spectra. Key contributions include a formal interpolation bound with , a latent representation , and a neural ray tracer that jointly models amplitude and phase along RF paths. Empirically, GRaF achieves state-of-the-art generalization to unseen scenes/layouts on simulated and real data, and improves downstream tasks such as AoA localization, while avoiding per-scene retraining required by prior NeRF-based RF methods.

Abstract

We present GRaF, Generalizable Radio-Frequency (RF) Radiance Fields, a framework that models RF signal propagation to synthesize spatial spectra at arbitrary transmitter or receiver locations, where each spectrum measures signal power across all surrounding directions at the receiver. Unlike state-of-the-art methods that adapt vanilla Neural Radiance Fields (NeRF) to the RF domain with scene-specific training, GRaF generalizes across scenes to synthesize spectra. To enable this, we prove an interpolation theory in the RF domain: the spatial spectrum from a transmitter can be approximated using spectra from geographically proximate transmitters. Building on this theory, GRaF comprises two components: (i) a geometry-aware Transformer encoder that captures spatial correlations from neighboring transmitters to learn a scene-independent latent RF radiance field, and (ii) a neural ray tracing algorithm that estimates spectrum reception at the receiver. Experimental results demonstrate that GRaF outperforms existing methods on single-scene benchmarks and achieves state-of-the-art performance on unseen scene layouts.

Paper Structure

This paper contains 34 sections, 1 theorem, 51 equations, 11 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Spatial Spectrum
  4. Vanilla NeRF for Spatial Spectrum Synthesis
  5. Related Work
  6. Methodology
  7. RF Spatial Spectrum Interpolation Theory
  8. Model Overview
  9. Latent RF Radiance Field
  10. Neural Ray Tracing Algorithm
  11. Experiments
  12. Experimental Setup
  13. Single-Scene Performance
  14. Generalization to Unseen Scenes
  15. Generalization to Unseen Layouts
  16. ...and 19 more sections

Key Result

Theorem 1

Let $\mathbf{SS}(\mathbf{P})$ denote the spatial spectrum of a transmitter at position $\mathbf{P} \in \mathbb{R}^3$, and let $\mathcal{N}_{L}(\mathbf{P}) = \{\mathbf{P}_i\}_{i=1}^{L}$ be its $L$ nearest neighbors with spectra $\mathbf{SS}_i$. Then the spectrum at $\mathbf{P}$ admits the approximati

Figures (11)

  • Figure 1: In the training scene, Radio-Frequency (RF) signals from each transmitter are measured across all surrounding directions by the receiver to form a spatial spectrum. Trained on this scene, GRaF synthesizes spectra for arbitrary transmitter locations in unseen scenes.
  • Figure 2: Antenna array; each blue square is an antenna.
  • Figure 3: Illustration of the spatial spectrum in 3D and 2D views.
  • Figure 4: Architecture of GRaF. Each voxel is represented by a feature $\mathbf{v}^s_i$, where $i$ indexes the $M$ rays, and $s$ denotes the voxel's position along the $i$-th ray. A neural-driven ray tracing algorithm computes the received signal power for each ray (TX: transmitter, RX: receiver).
  • Figure 5: Top-view visualization of two room layouts.
  • ...and 6 more figures

Theorems & Definitions (1)

  • Theorem 1: Spectrum Interpolation