GSpaRC: Gaussian Splatting for Real-time Reconstruction of RF Channels

TL;DR

<3-5 sentence high-level summary> GSpaRC introduces a real-time RF channel reconstruction framework based on Gaussian splatting, modeling the environment with a compact set of 3D Gaussian primitives augmented by physics-informed features and a lightweight per-Gaussian MLP. It employs an equirectangular hemispherical projection and a CUDA-accelerated renderer to achieve sub-millisecond inference within 1 ms TTIs, enabling predictive CSI reconstruction from sparse RF measurements. Compared to NeRF2 and WRFGS+, GSpaRC delivers comparable reconstruction fidelity with orders-of-magnitude reductions in training and inference time. The approach demonstrates strong potential for reducing pilot overhead and powering low-latency, beamforming-enabled wireless systems in 5G and beyond, with applicability to digital twins and mobility-aware networking.

Abstract

Channel state information (CSI) is essential for adaptive beamforming and maintaining robust links in wireless communication systems. However, acquiring CSI incurs significant overhead, consuming up to 25\% of spectrum resources in 5G networks due to frequent pilot transmissions at sub-millisecond intervals. Recent approaches aim to reduce this burden by reconstructing CSI from spatiotemporal RF measurements, such as signal strength and direction-of-arrival. While effective in offline settings, these methods often suffer from inference latencies in the 5--100~ms range, making them impractical for real-time systems. We present GSpaRC: Gaussian Splatting for Real-time Reconstruction of RF Channels, the first algorithm to break the 1 ms latency barrier while maintaining high accuracy. GSpaRC represents the RF environment using a compact set of 3D Gaussian primitives, each parameterized by a lightweight neural model augmented with physics-informed features such as distance-based attenuation. Unlike traditional vision-based splatting pipelines, GSpaRC is tailored for RF reception: it employs an equirectangular projection onto a hemispherical surface centered at the receiver to reflect omnidirectional antenna behavior. A custom CUDA pipeline enables fully parallelized directional sorting, splatting, and rendering across frequency and spatial dimensions. Evaluated on multiple RF datasets, GSpaRC achieves similar CSI reconstruction fidelity to recent state-of-the-art methods while reducing training and inference time by over an order of magnitude. By trading modest GPU computation for a substantial reduction in pilot overhead, GSpaRC enables scalable, low-latency channel estimation suitable for deployment in 5G and future wireless systems. The code is available here: \href{https://github.com/Nbhavyasai/GSpaRC-WirelessGaussianSplatting.git}{GSpaRC}.

Figures (8)

• Figure 1: Reconstruction of signal intensity over a hemisphere at a fixed receiver for different transmitter positions. GSpaRC provides high accuracy at sub-millisecond rendering times.
• Figure 2: GSpaRC pipeline overview: (a) Uniform initialization of 3D Gaussians across the scene; (b) Physics-informed signal modeling using per-Gaussian MLPs and path loss; (c) Equirectangular projection of Gaussians onto a 2D hemisphere from the receiver’s view; (d) Parallelized rasterization with depth sorting and $\alpha$-compositing via CUDA; (e) Synthesized spatial spectrum prediction compared with ground truth. The pipeline supports end-to-end differentiable optimization.
• Figure 3: Comparison of GSpaRC, NeRF2, and WRF-GS+ across SSIM (↑), PSNR (↑), and MSE (↓).
• Figure 4: The left panel shows the ray-tracing paths generated in Sionna for the indoor conference room scene, illustrating line-of-sight and reflection paths between the transmitter (red) and receiver (green). The right panels compare the ground-truth and GSpaRC-predicted spectra, shown in both grayscale (top) and polar coordinates (bottom), highlighting the strong spatial correspondence between simulated and reconstructed field distributions.
• Figure 5: Comparison of GSpaRC, NeRF2, and WRF-GS+ across SSIM (↑), PSNR (↑), and MSE (↓).