GSBF: Gaussian Splatting for Environment-Aware Beamforming

Summary of the paper GSBF: Gaussian Splatting for Environment-Aware Beamforming

This paper proposes a novel framework called GSBF (Gaussian Splatting for environment-aware beamforming) that leverages 3D Gaussian splatting and multi-modal data to synthesize MIMO beamforming vectors without requiring online instantaneous channel state information (CSI).

Motivation and Problem Statement

The authors note that "Beamforming plays a key role in multiple-input-multiple-output (MIMO) communication systems. However, conventional beamforming design normally requires accurate instantaneous channel state information (CSI) and iterative optimization, which incur substantial pilot overhead and computational complexity. While deep learning has been introduced to reduce online iterative overhead, these methods still rely on designed pilots. Channel Knowledge Maps (CKM) were proposed but are hindered by the prohibitive overhead of site-specific wireless measurements and storage, while fixed codebook resolutions inevitably introduce performance loss. Multi-modal approaches relying on line-of-sight sensors are prone to outages in non-line-of-sight (NLoS) scenarios. The paper also criticizes prior implicit neural representations since the implicit environment representations within neurons lack spatial causality and physical consistency."

The paper makes three main contributions: "1) We propose an online CSI-free and differentiable pipeline that deterministically maps a persistent 3D Gaussian scene representation to beamforming vectors; 2) within this pipeline, we design a bidirectional spherical Gaussian (Bi-SG) kernel that explicitly models the directional dependence of multipath effects while enforcing physical reciprocity; and 3) we further propose a two-sided electromagnetic rasterization and synthesis strategy that facilitates the pipeline by aggregating complex-valued angular propagators into constant-modulus beamformers through an over-complete dictionary."

System Model

The paper considers an indoor scenario where the user is moving and the AP is static. The AP is equipped with a uniform planar array (UPA), with Nx and Ny elements along the x and y axes, respectively. The received signal is expressed as y t = √P t h t f t x t + n t, where the SNR is given by SNR t = P t h t f t2 / (N0·BW) and the spectral efficiency is R t = log2(1 + SNR t). Since the AP has a single radio frequency (RF) chain connected to an array of phase-shifters, thus the beamforming vector f t has constant-modulus constraints. The objective is to maximize SE with these constraints, which renders the problem non-convex.

Preliminaries on 3DGS

The paper reviews that 3DGS is a recently proposed explicit scene representation approach for high-quality and real-time novel view synthesis. It models a three-dimensional scene as a collection of explicit, anisotropic Gaussian primitives. Unlike implicit neural representations, 3DGS offers a persistent geometric backbone that is both differentiable and computationally efficient. Each primitive is described by its spatial mean μ i and covariance Σ i = R i S i S i T R i T. The paper notes that "Recent 3DGS extensions to wireless radiation fields interpret rendered images as power angular spectra (PAS) or channel gain maps. However, these real-valued radiometric reconstructions cannot directly support complex-valued beamforming design... Moreover, their tight coupling between array dimensions and rendered PAS prevents environment-array decoupling. These drawbacks motivate our GSBF framework for direct beamforming generation from a persistent Gaussian representation."

GSBF Framework

The beamforming synthesis is framed as: "The synthesis of an analog beamforming vector f can be interpreted as the coherent reconstruction of a desired wavefront from the angular domain. Mathematically, f is a weighted superposition of steering vectors a q spanning Q directions, which can be calculated as f = Σ q=1 Q b q a q, where b q denotes the angular propagator in direction q. They employ an over-complete basis dictionary where Q ≥ N."

Bi-SG Kernel Design: Since standard 3D Gaussians lack the intrinsic mechanisms to describe complex electromagnetic interactions, the authors design a bidirectional spherical Gaussian kernel. "For a signal incident from direction ω in and scattering towards direction ω out, the Bi-SG kernel parameterizes the scattering response of the i-th Gaussian primitive S i as S i(ω in, ω out) = γ i + Σ k=1 K a k,i exp[λ k,i(κ k,i − 1)]," where γ i and a k,i are learnable complex-valued albedo and lobe amplitudes, v k,i is the unit-normalized learnable axis, and λ k,i is the sharpness. The alignment metric κ blends collinear propagation and redirected-scattering terms with a sigmoid-based collinearity weight. Crucially, this formulation in (7) ensures S i(ω in, ω out) = S i(−ω out, −ω in) under link reversal, guaranteeing consistency with reciprocal wave propagation.

Panoramic Projection: The framework "adopts an equirectangular projection to first project the 3D Gaussian primitives onto a panoramic angular map with W × H resolution, and then crop the result to obtain the angular propagator map corresponding to the antenna’s forward hemisphere. The 3D covariance is propagated to the 2D plane via Σ'i = J i W Σ i W T J i T, with the Jacobian J i provided. This panoramic projection strategy ensures the spatial causality of the environment is preserved in the angular domain."

Electromagnetic Rasterization: The process is divided into two parts. First, source-to-primitive transmittance computes a factor τ i UE that quantifies the residual signal reaching the i-th Gaussian primitive, accounting for occluders via front-to-back cumulative products: τ i UE = Π j∈R i (1 − δ ij UE). Second, coherent propagator accumulation synthesizes the propagator map at the AP side. The contribution is b i(u, v) = o i τ i UE τ i AP(u, v) S i(ω in, ω out) G′(q′; q′ i, Σ i AP), where τ i AP accounts for AP-side transmittance of overlapping primitives. The aggregate propagator is b u,v = Σ i∈S u,v b i(u, v), producing B t = b u,v ∈ C H×W′ for the forward-hemisphere angular space.

Beamforming Generation: The angular propagator map is vectorized into b t = vec(B t), then synthesized as f̃ t = A b t using the over-complete array-manifold dictionary A, where steering vectors are defined with phase shifts Ψ m,n = mΦ x + nΦ y. Since the beamforming vector must satisfy the constant-modulus requirement for each antenna element, they project onto the feasible domain: f̂ t i = f̃ t i / (√N f̃ t i).

The optimization loss is L = D t + λ cm L cm, where D t is the normalized beamforming power-gain gap and L cm penalizes amplitude variation of the unnormalized vector. Optimized parameters include Gaussian spatial parameters μ i, R i, S i and the Bi-SG kernel parameters v k,i, γ i, o i, a k,i, λ k,i, together with density control. The straight-through estimator (STE) is used to mitigate the gradient mismatch from projection.

Simulation Setup and Results

A synthetic dataset was created based on the factory scene constructed from open-source components, using scene construction in Blender and ray-tracing in Sionna. The setup used central frequency f c = 28 GHz, with bandwidth BW = 100 MHz, noise power density N0 = −174 dBm/Hz, and transmit power P T = 10 dBm. During training, the model gets regional LiDAR data, the AP pose, the user position, and the corresponding channel, but at evaluation, only the AP pose and user position are required. The rendered angular map has H = W′ = 45, with K = 4 lobes, ν = 80, ξ = 0.95, λ cm = 0.1, trained on 5600 samples and tested on 1400 samples.

The results in Table I show GSBF achieves mean/median SE (bps/Hz) of 7.798/7.932 for 4×4, 8.874/9.051 for 6×6, and 9.625/9.812 for 8×8 arrays. This consistently outperforms EBA (7.203/7.227, 8.238/8.216, 9.066/9.071) and LocBF (4.485/3.824, 3.771/2.671, 3.648/2.415), while approaching the fully-digital DBF upper bound (8.281/8.261, 9.490/9.437, 10.345/10.252). The authors state: GSBF enables fine-grained beam steering without the quantization losses inherent in the discrete codebooks used by EBA, and can effectively capture the complex scattering patterns while LocBF degrades as the array size increases due to its inability to account for NLoS multi-path components.

For latency, GSBF also achieves an average inference latency of 24.2 ms, 20.9 ms, and 23.5 ms for 4×4, 6×6, and 8×8 arrays, respectively, whereas "EBA incurs substantial physical overhead from sequential beam sweeping: in practical mmWave systems with 8 synchronization signal blocks per 20 ms period, the alignment latency is approximately 40.0 ms, 100.0 ms, and 160.0 ms. GSBF's latency is stable because the rasterization complexity depends primarily on the number of Gaussian primitives and the panorama resolution, both largely independent of the array dimensions."

From the ECDF curves, the authors observe the dataset mainly includes two distinct propagation regimes, which are inherently rooted in the heterogeneous propagation characteristics of the indoor factory environment, with NLoS paths dominating. At high SE values, GSBF outperforms EBA and LocBF across all considered array sizes while remaining close to DBF. The remaining gap to DBF results from both the constant-modulus analog beamforming constraint and the approximation error of the learned GSBF representation.

Conclusion

The paper concludes: "This paper proposed GSBF, a framework that synthesizes grid-free beamforming vectors directly from multi-modal data without online CSI. By designing panoramic projection, two-sided electromagnetic rasterization, and a reciprocity-preserving Bi-SG environment surrogate, GSBF renders an angular propagator map that is mapped to constant-modulus beams via an over-complete array-manifold dictionary. Simulations confirm that GSBF approaches the digital beamforming upper bound while outperforming baselines with millisecond-scale inference, establishing a new paradigm for environment-aware beamforming through the fusion of physical modeling and data-driven learning."

Improvement 1: Replace implicit neural scene representations with explicit, differentiable 3D Gaussian primitives that persist across time and tasks.

The improved AI system can maintain a persistent, spatially-causal map of an environment and instantly synthesize task-specific outputs (e.g., beamforming vectors, channel maps) for arbitrary query positions, without per-position retraining or online pilot measurements.

Improvement 2: Embed physics-aware, bidirectional spherical Gaussian (Bi-SG) kernels into the scene representation to model directional, complex-valued multipath scattering.

The improved AI system can enforce electromagnetic reciprocity and directional dependence, yielding physically consistent predictions in non-line-of-sight (NLoS) conditions, and generalizing to unseen array sizes or user locations with higher accuracy than conventional neural models.

Improvement 3: Introduce two-sided electromagnetic rasterization with cumulative transmittance from both source and receiver perspectives.

The improved AI system can reason about occlusions, shadowing, and coherent wave accumulation in complex environments, directly producing complex-valued angular propagator maps that are useful for beamforming, radio mapping, and sensing tasks.

Improvement 4: Use an over-complete array-manifold dictionary to project high-dimensional angular outputs into feasible, constant-modulus control signals.

The improved AI system can generate hardware-compliant outputs under strict non-convex constraints (e.g., analog phase-shifters) in a differentiable manner, avoiding codebook quantization losses and enabling fine-grained control.

Improvement 5: Decouple environment representation from antenna array geometry via panoramic equirectangular projection.

The improved AI system can reuse a single environment model for multiple array configurations (4×4, 6×6, 8×8) and maintain constant millisecond-scale inference latency, independent of array dimensions.

Improvement 6: Train end-to-end with multi-modal data (LiDAR, poses, channels) using a straight-through estimator to handle non-differentiable projection steps.

The improved AI system can fuse geometric sensor data with wireless channel observations during training, and at inference require only coarse position information (no CSI), reducing pilot overhead and system complexity.

Improvement 7: Optimize a hybrid loss combining beamforming power-gain gap with an amplitude-regularization term.

The improved AI system can balance spectral-efficiency maximization against hardware feasibility, approaching the digital beamforming upper bound while operating under analog constraints.

Improvement 8: Leverage explicit Gaussian density control and learnable covariance to adapt scene complexity dynamically.

The improved AI system can automatically concentrate representational capacity on electromagnetically important regions, improving sample efficiency and prediction accuracy in heterogeneous indoor industrial environments.

Overall, the improved AI system can perform environment-aware, CSI-free beamforming in real time, achieve near-optimal spectral efficiency, support multiple array geometries from one shared scene model, and generalize to NLoS scenarios—all with millisecond-level inference and dramatically reduced training/online overhead compared to conventional iterative optimization or deep-learning methods.