Technical Summary: RadCloudSplat

Problem Statement

Constructing accurate radiomaps—spatial distributions of wireless signal strength (RSS)—is critical for network optimization, resource allocation, and autonomous driving. However, acquiring dense signal measurements in large-scale outdoor environments is prohibitively expensive and labor-intensive, typically relying on sparse data collected via drive tests. Traditional reconstruction methods face significant limitations:

• Interpolation-based methods (e.g., Inverse Distance Weighting, Kriging) fail to capture the complex physics of wireless propagation in outdoor settings.
• Ray Tracing requires precise 3D CAD models that often diverge from real-world environments and are computationally heavy.
• Learning-based methods (e.g., VAEs, GANs) generally require dense training data and lack explicit physical modeling.
• Neural Radiance Fields (NeRF) have shown promise but suffer from slow synthesis speeds and high computational costs, making them impractical for large-scale deployment.
• Existing 3D Gaussian Splatting (3DGS) adaptations (e.g., WRF-GS) are often limited to indoor scenarios, rely on simplified channel assumptions, or lack the ability to utilize environmental point cloud priors effectively.

The core challenge is to achieve efficient, accurate, and robust radiomap extrapolation from sparse RSS measurements in complex outdoor environments by leveraging environmental geometry without relying on proprietary antenna parameters or dense ground truth data.

Methodology: RadCloudSplat Framework

The authors propose RadCloudSplat, a framework that extends 3D Gaussian Splatting (3DGS) to the radio frequency domain. It models environmental scatterers (buildings, walls) as anisotropic 3D Gaussians to capture multi-path propagation characteristics. The framework consists of four key components:

1. Relaxed-Mean (RM) Reparameterization

To address the challenge of processing noisy, dense LiDAR point clouds, the authors introduce a Relaxed-Mean (RM) scheme.

• Goal: Select key virtual scatterers from raw point clouds ($\hat{P}_{raw}$) to serve as the means ($\mu_{rad}$) for 3D Gaussians.
• Mechanism: Instead of treating scatterer positions as fixed points from the cloud, the model learns a trainable selection matrix $T$ and a bias term $B$. The positions are reparameterized as $\mu_{rad} = T\hat{P}_{raw} + B$.
• Optimization: The selection matrix $T$ is constrained to a simplex space with a Maximum Element Constraint (MEC) to ensure sparsity (selecting distinct points) and row-wise dominance. The bias $B$ allows the model to refine positions to mitigate noise and better align with signal propagation physics. This transforms a discrete selection problem into a differentiable optimization task.

2. Camera-Free 3DGS Projection

Unlike traditional 3DGS which maps 3D scenes to 2D camera views, RadCloudSplat maps 3D scatterers to a 2D beamspace (RSS patterns across beam directions).

• Challenge: Wireless signals lack intrinsic/extrinsic camera parameters, and operator-specific antenna patterns are often proprietary.
• Solution: A camera-free projection uses two structured Multi-Layer Perceptrons (MLPs) to learn the mapping from 3D scatterer attributes (mean $\mu_{rad}$ and covariance $\Sigma_{rad}$) to 2D beamspace Gaussian parameters (mean $x$ and precision $\Lambda$). This avoids reliance on unknown physical antenna models.

3. Electromagnetic Splatting

The rendering process synthesizes RSS values by simulating multi-path propagation:

• Scatterers as Virtual Transmitters: Each 3D Gaussian scatterer is treated as a secondary source that re-transmits signals with specific attenuation and phase shifts.
• View-Dependent Encoding: Spherical Harmonics (SH) coefficients encode the direction-dependent contribution of each scatterer.
• Depth Sorting & Blending: Scatterers are sorted by depth relative to the target location. The final RSS is computed by aggregating the complex contributions of all sorted scatterers, accounting for accumulated attenuation and phase shifts, effectively simulating the superposition of multi-path signals.

4. Training and Extrapolation Strategy

• Loss Function: A composite loss combines Mean Squared Error (MSE) for accuracy and Total Variation (TV) loss to ensure spatial smoothness and coherence.
• Recursive Fine-Tuning: To handle the sparsity of measurements, the authors propose a recursive scheme. The model is first trained on known sparse grids. It then predicts values for boundary-adjacent unknown grids; these predictions are added to the training set, and the model is retrained. This iterative process progressively expands the coverage until the entire area is extrapolated.

Key Contributions

1. First Extension of 3DGS to Radio Frequencies: The paper presents the first framework to apply 3D Gaussian Splatting to outdoor radiomap reconstruction using beamspace RSS data, moving beyond visible light and simplified indoor models.
2. Relaxed-Mean (RM) Reparameterization: A novel method to extract key virtual scatterers from noisy, dense point clouds by learning a differentiable selection and refinement process, enabling the model to utilize environmental priors effectively.
3. Camera-Free Projection: A learning-based projection mechanism that maps 3D scatterers to 2D beamspace without requiring proprietary antenna gain or beamforming patterns, enhancing robustness in real-world deployments.
4. Extrapolation-Enhancing Techniques: The integration of a ternary regularized loss function and a recursive fine-tuning scheme to achieve robust generalization in regions with sparse or no measurements.

Experimental Results

The framework was evaluated on both synthetic data (City A) and real-world drive-test data (City B).

• Synthetic Data (City A): RadCloudSplat achieved a Mean Absolute Error (MAE) of 7.564 dB, outperforming baselines including Kriging (11.307 dB), VAE (10.306 dB), and NeRF-based methods (NeRF2: 9.867 dB, NeWRF: 10.051 dB). It also demonstrated superior speed, with an inference time of 0.004 seconds per sample, approximately 5x faster than NeRF variants.
• Real-World Data (City B): On real-world data with inherent sensor noise and GPS inaccuracies, RadCloudSplat achieved an MAE of 7.035 dB, outperforming Ray Tracing (13.108 dB), Kriging (9.656 dB), and NeRF2 (7.313 dB).
• Ablation Studies:
  • Removing the recursive fine-tuning strategy increased the MAE to 7.967 dB (synthetic) and 7.121 dB (real-world), confirming its importance for extrapolation.
  • Replacing the learned camera-free projection with a physical projection (assuming known antenna patterns) significantly degraded performance (MAE 12.218 dB), validating the necessity of the learned projection for real-world scenarios where such parameters are unknown.
  • The RM scheme proved critical; random selection of scatterers or fixed initialization without the bias term resulted in significantly higher errors.

Significance and Claims

The paper claims that RadCloudSplat solidifies the credibility of 3DGS-based models for real-world wireless network deployment. By successfully integrating environmental point cloud priors with radio propagation physics, the framework addresses the computational inefficiencies of NeRF and the physical limitations of traditional interpolation.

The authors emphasize that the framework enables robust and spatially-aware modeling of wireless signal radiation patterns from sparse measurements. The ability to learn scatterer positions and projection parameters directly from data, without relying on proprietary operator information, makes the approach particularly suitable for practical MIMO-based deployments and next-generation network optimization tasks. The work demonstrates that advanced 3D modeling techniques can be effectively adapted to the radio domain to solve critical challenges in signal extrapolation.