iComAI Lab
INTEGRATED COMMUNICATIONS AND AI LAB
Point2Radio: A Foundation Model for Cross-Scene Radio Fields from Material-Aware Point Clouds Chaozheng Wen†1 , Chenghong Bian†1 , Hongze Chen1 , Jun Zhang* 1 1
arXiv:2607.28994v1 [cs.NI] 31 Jul 2026
†
Hong Kong University of Science and Technology Equal Contributions
*
Corresponding Authors
High-fidelity radio fields are typically simulated for every scene–transmitter configuration or fitted separately to each scene, failing to exploit propagation structures shared across environments. We present Point2Radio, a foundation model that learns a transferable propagation prior from multiple environments. Given a material-aware point cloud and a transmitter (TX) setting, a common encoder produces a TX-conditioned scene representation that can be queried at arbitrary receiver (RX) locations. Task-specific query decoders map this representation to different radio quantities, e.g., three-dimensional (3D) path-gain (PG) fields and power angular spectra (PAS). At inference for a new scene, the model uses only a material-aware point cloud and transceiver queries, running in milliseconds on a single GPU without meshes or explicit path tracing. We evaluate PG prediction on a scene-disjoint split of a 337-scene corpus containing 86,272 TX-conditioned fields. Point2Radio achieves 0.871 dB mean absolute error (MAE), reducing error by 76.7% relative to a same-split UNet-style baseline. The same encoder also supports PAS prediction via a task-specific decoder. Experiments further show that light target-scene fine-tuning improves adaptation to a specific environment. Code: https://github.com/wenchaozheng/Point2Radio Correspondence to: Jun Zhang ([email protected])
1
Introduction
Modeling radio fields in 3D environments is increasingly important for low-altitude aerial networks, embodied AI, connected autonomous vehicles, and multi-robot coordination Zeng et al. (2024). It can guide coverageaware planning and reliable multi-agent communication, while RF observations offer embodied agents an additional sensing modality Guo et al. (2025). These applications all benefit from knowing how a transmitted signal arrives across a scene. Radio field prediction is a long-standing problem, yet accurate and transferable solutions remain scarce. Radio waves undergo reflection, transmission, scattering, and diffraction on complex surfaces and materials. These interactions rarely admit accurate closed-form solutions, so high-fidelity fields are usually obtained only through expensive numerical simulation or dense on-site measurement. Recent progress in AI and foundation models has transformed vision Carion et al. (2025) and language Achiam et al. (2023), but has transferred less cleanly to radio-field modeling. Radio labels are often scarce, costly to collect, and inconsistent across devices, frequencies, and sites. As a result, learning systems in this area still struggle to acquire broad, reusable propagation knowledge. Existing approaches attack the problem from several directions. Classical empirical models Sarkar et al. (2003) are inexpensive but cannot resolve detailed spatial variation. Deterministic ray tracing instead constructs propagation paths and applies parameterized interaction models. Increasing the ray density, path depth, and supported mechanisms improves fidelity, but computational cost grows rapidly, and the search must be truncated and repeated for every new scene and TX configuration Hoydis et al. (2023). This fidelity– computation trade-off motivates learning recurring propagation patterns from expensive offline simulation and reusing that experience across configurations, rather than recomputing paths at inference.
1
Scene
ROOM1 (z=3)
ROOM2 (z=15)
ROOM3 (z=7)
Figure 1 Visualization of cross-scene generalization on three completely unseen indoor rooms. Left: 3D point cloud with the selected height slice marked by the cutting plane. Remaining columns: ground-truth (GT) PG map and predictions from zero-shot (ZS) Point2Radio, RadioUNet, 3D U-Net, NeRF2 , and WRF-GS+. Slice-wise MAE↓ / peak signal-to-noise ratio (PSNR)↑ / structural similarity (SSIM)↑ are reported under each map. Since NeRF2 and WRF-GS+ are per-scene methods without cross-scene transfer, we train each with a single randomly chosen TX sample (N=1) for a fair low-data comparison against ZS baselines.
Learning-based methods currently occupy two main operating points. Amortized cross-scene predictors such as RadioUNet Levie et al. (2021) are feedforward and transferable, but typically map two-dimensional (2D) layouts to fixed path-loss rasters and thus omit the 3D geometry and material interactions that govern multipath propagation. Scene-specific neural RF representations take the opposite trade-off: methods such as NeRF2 Zhao et al. (2023) and WRF-GS Wen et al. (2025) recover detailed fields within one environment, but require target-scene measurements or optimization and do not reuse a single model across scenes. Existing work thus tends either to amortize across scenes while compressing geometry into 2D, or to retain 3D fidelity by refitting each environment. These limitations invite a natural question: can a model learn transferable priors about how radio signals interact with 3D scenes, rather than treating each environment as a fresh computation or fit? If such priors are captured once, cross-scene reuse becomes possible, and different decoders can be trained on the same representation for different downstream radio quantities. We therefore cast radio fields prediction in the usual foundation model form of a shared backbone with task-specific heads: H = E(R, t),
ŷ = D(H, q).
(1)
Here R denotes a structured 3D scene representation that carries geometry and electromagnetic material information, E encodes R together with the TX setting t into a scene code H, and a task-specific decoder D answers query q. In principle, any radio quantity queried at an RX from (H, q) fits this interface. We propose Point2Radio as one realization of this paradigm. We first represent each scene as a material-aware point cloud. Structured hierarchical tokenization then aggregates scene points into a compact token set. These tokens are fed with the TX into Cross–Self–Cross (CSC) attention, which yields TX-conditioned latent codes. At inference, each query gathers nearby codes by kNN and a task-specific head predicts the radio quantity. The shared encoder transfers ZS to unseen scenes, and can be further adapted by light decoder fine-tuning when target scene labels are available. This paper validates the design on PG prediction for dense spatial fields (Figure 1) and PAS prediction for directional spectra (Figure 6). Our main contributions are: 1. We present Point2Radio, a foundation model for cross-scene radio fields that predicts RX-centric quantities from a material-aware point cloud and a TX setting. Hierarchical tokenization and CSC attention encode TX–scene interactions into transferable latent codes. 2
2. We show that a PG-pretrained encoder can be reused across tasks by attaching task-specific query decoders while keeping the backbone frozen, enabling fast transfer from PG to PAS. Light residual adaptation further improves performance on a target scene. 3. We introduce PRISM, a Procedural Radio–Image Scene Multimodality dataset, which provides aligned wireless and vision annotations on procedurally generated indoor scenes. This paper uses the wireless splits PRISM-PG, with dense PG volumes for 337 scenes, and PRISM-PAS for directional experiments. 4. Extensive experiments on PG and PAS prediction show that Point2Radio is both accurate and efficient, enabling millisecond single-GPU inference with simple inputs and low compute cost.
2
Related Work
2.1
Per-scene radio field prediction.
Radio propagation depends strongly on the surrounding geometry and materials, which has motivated transferring neural scene representations from computer vision to per-scene radio field modeling. Neural radiance fields (NeRF) Mildenhall et al. (2021) represent a scene as a continuous field optimized from sparse observations. This paradigm has been adapted to wireless channel prediction within a site Lu et al. (2024); Wang et al. (2026a). More recently, 3D Gaussian splatting (3DGS) Kerbl et al. (2023) replaces implicit MLPs with explicit Gaussian primitives for faster high-quality reconstruction. Subsequent radio works adopt 3DGS for spatial spectrum reconstruction, with physical propagation priors or visual priors Wen et al. (2026a); Zhang et al. (2026a). Closely related extensions keep a scene-level Gaussian representation while broadening the radio target, including cross-frequency radiation fields with shared geometry and frequency-adaptive RF attributes Wang et al. (2026b), and geometry-conditioned delay–beam priors for high-mobility channel estimation Zhang et al. (2026b). A common bottleneck is that many of these reconstructions specialize to a fixed TX or a fixed RX setting. Subsequent methods improve TX–RX flexibility inside a reconstructed scene. RFCanvas Chen et al. (2024) adapts the model with visual priors and few-shot radio measurements. URF-GS Wen et al. (2026b) uses physics-informed inverse rendering to separate emission from environment properties and synthesize new TX–RX pairs. RayProNet learns a neural point-field surrogate that can be queried at new TX/RX locations after scene-specific training Cao and Peng (2024). Even with such within-site flexibility, each new environment still requires its own optimization. In contrast, Point2Radio learns a transferable scene representation for scene-disjoint prediction, without target-scene fitting.
2.2
Cross-scene radio field prediction.
Differentiable ray-tracing methods take another path to new environments. Once a scene is available, they Orekondy et al. (2023); Chen et al. (2025) can synthesize channels for new TX–RX links through an explicit 3D propagation loop. Radio digital twins An et al. (2025) similarly couple mapped geometry with learned electromagnetic materials in a differentiable tracer. These approaches have a degree of scene-level generalization ability, but the transfer is geometry-conditioned. High-fidelity meshes and material information are often difficult to obtain at scale, and new sites commonly still need sparse measurements to calibrate materials or interaction parameters. Cross-scene radio map networks instead pursue a feedforward manner without specific scene fitting, predicting path-loss coverage from TX-conditioned layout rasters Lee et al. (2023). Generative models further extend this paradigm to dynamic or multi-height radio-map tensors Wang et al. (2026c). While they avoid per-scene ray tracing, their inputs remain occupancy-style maps and their outputs are discretized rasters, so fine-grained 3D surface geometry and materials are easy to lose. Another line of work trains foundation models on wireless measurements. WiFo Liu et al. (2025) pretrains a masked model to complete channel tensors from partial observations. MUSE-FM Zheng et al. (2026) learns a shared backbone for several downstream radio tasks and conditions on a 2D top-down layout map. These designs remain limited in capturing fine-grained 3D geometry and materials for dense spatial field queries. Relative to this cross-scene line, Point2Radio takes a material-aware point cloud rather than a layout raster, measurement tensor, or calibrated mesh. It predicts
3
RADIO TRANSPORT ENCODER (E) Hierarchical Tokenization
Cross-Attn
...
Position
Scene Tokens
×N
kNN RX
(q)
TX-Conditioned Scene Tokens (H)
Embedding
PG Decoder (X × Y × Z)
PAS Decoder
...
Positions + Normals + Materials
...
Scene Point Clouds TXTXSettings Settings
×N
Cross-Attn
CSC Attention
×N
...
LOCAL QUERY DECODER (D)
Self-Attn
INPUTS (R,t)
TX Token
(� × �)
Figure 2 Overview of Point2Radio. Inputs are a material-aware point cloud and a TX setting. The encoder tokenizes
the scene and applies CSC attention to form TX-conditioned latent codes. The local query decoder aggregates nearby features for task-specific PG or PAS prediction.
radio quantities at dense 3D query points in one feedforward pass without online ray tracing or target-scene radio fitting.
3
Method
In this section, we present Point2Radio, which realizes the encoder–decoder process in Eq. (1), as depicted in Figure 2. Given a material-aware point cloud and a TX setting, we first construct the scene representation R by attaching electromagnetic attributes and an explicit TX emission point. The encoder E then hierarchically tokenizes the scene and fuses tokens with the TX through CSC attention, producing TX-conditioned latent codes H. Finally, the decoder D applies the same local kNN query mechanism at the RX location for both PG and PAS, and differs mainly in the task heads that map the aggregated features to scalar PG or directional PAS. Finally, we attach a lightweight residual head for fine-tuning, enabling fast adaptation to a new target scene.
3.1
Input Representation
The pipeline begins by constructing the scene representation for the encoder. We represent each indoor environment as a material-aware point cloud. Compared with structured geometry such as meshes or volumetric grids, point clouds are more flexible and more accessible. The sampling density can be adapted to scene complexity, enabling a balance between efficiency and accuracy. Moreover, the absence of explicit connectivity among points simplifies subsequent neural processing. In practice, each scene is discretized into M surface points. Every point is assigned a geometric and electromagnetic feature: M R = {(xi , ni , ϵi , σi , si , χi , ei )}i=1 , (2) where xi and ni denote the position and surface normal. Relative permittivity ϵi and conductivity σi follow the ITU material model itu (2023). For numerical stability, and to avoid domination by large values such as metal conductivity (≈107 S/m), we store σi as log10 σi . A scattering coefficient si and a cross-polarization discrimination coefficient χi are further attached according to the material class. To distinguish scene points from the TX, we add a binary emission indicator that is zero on every point in R and one on the TX. The TX setting t is then formed from its location together with this emission channel. The pair (R, t) is then passed to the radio transport encoder E, which maps it to TX-conditioned latent codes H.
3.2
Radio Transport Encoder
Given (R, t), the encoder E produces a reusable TX-conditioned scene code H. Explicit path tracing is costly at inference and does not learn a transferable prior. Following Xu et al. (2026), multi-bounce transport can be viewed as repeated scene-wide mixing of an emission under a transport operator, and attention offers a related all-to-all aggregation as a feedforward surrogate. The same operator view carries over to indoor radio, where a TX emission reaches RXs through multipath interactions with material surfaces, including reflection, 4
diffraction, and transmission. We therefore use attention for TX-conditioned propagation encoding rather than online path tracing. To this end, we design E to model propagation between the TX and scene points through hierarchical tokenization followed by CSC attention, as shown in Figure 2. Hierarchical tokenization. Attending over every point in R is too expensive, since self-attention scales quadratically with the number of tokens. We compress R into fewer tokens while retaining local geometry and material cues. A shared MLP first lifts per-point features. We then apply two downsampling stages. Each stage selects centers by farthest-point sampling (FPS), gathers k nearest neighbors for every center, and t aggregates neighbor features with an MLP followed by max-pooling. This yields Nt tokens S = {sj }N j=1 of t width d, located at the retained centers {xj }N j=1 . The TX setting t is encoded by a separate MLP into etx . 3 We write xtx ∈ R for the TX location in t, which later provides relative geometry for TX–scene and TX–RX terms. TX-conditioned CSC attention. CSC models TX–scene interactions and stores the result in the scene tokens for later RX decoding. As shown in the Figure 2, we stack TX-to-scene cross-attention, scene self-attention, and another TX-to-scene cross-attention, each repeated N times. The first cross-attention brings TX information into the scene tokens, self-attention exchanges information among tokens, and the second cross-attention updates the tokens again with the TX. Each TX-to-scene block uses vector cross-attention with relative geometry: ∆xj = xtx − xj , Pj = γ(∆xj ), (3) qj = Wq sj ,
k = Wk etx ,
v = Wv k, Aj = softmax ψ(k − qj + Pj ) , s̃j = Wo Aj ⊙ (v + Pj ) .
(4) (5) (6)
Here γ and ψ are small MLPs, Wq , Wk , Wv , Wo are learned projections, and ⊙ is element-wise multiplication. Nt t The output H = {sj }N j=1 remains anchored at {xj }j=1 .
3.3
Local Query Decoder
After CSC, each retained token stores a TX-conditioned code at a surface location. Interpreting these tokens as virtual scatterers, we decode each RX by aggregating the nearest codes around the query. Since CSC self-attention has already propagated long-range scene context across tokens, this local neighborhood contains both local geometric cues and non-local propagation information, making local aggregation a principled readout. Local query aggregation. Given an RX location q, we first build a query embedding with an MLP over the concatenated features q, q − xtx , ∥q − xtx ∥ , (7) which encode the RX position together with its displacement and distance to the TX. Next, we select the K codes in H whose anchors are nearest to q. Finally, the query embedding attends to these neighbor codes with the same vector cross-attention as in CSC, using the relative displacements from q to the neighbor anchors as the positional term. The updated embedding is the RX latent fed to the task head. Task heads. PG and PAS share the encoder and the local readout, and differ only in the task head applied to the RX latent. For PG, the query is an arbitrary 3D location q. An MLP maps the RX latent to a scalar PG ŷpg (q). For PAS, we keep the same RX latent and further query a direction of azimuth ϕ and elevation θ. Let u(ϕ, θ) be the corresponding unit vector. We form Fourier features ξ(u) = u, {sin(2k πu), cos(2k πu)}L−1 (8) k=0 , embed ξ(u) with an MLP, concatenate the result with the RX latent, and map the fused features to directional power ŷpas (q, ϕ, θ). After PG pretraining, we freeze the spatial encoder and train only this PAS head, transferring the learned propagation prior to directional spectra. 5
(a)
(b)
(c)
(d)
Figure 3 PRISM dataset overview. Examples of: (a) floor plan; (b) untextured meshes for a floor plan; (c) textured
mesh objects; (d) meshes colored by material class, with the legend on the right.
Target-scene residual adaptation. ZS inference already transfers a shared propagation prior to unseen rooms, but a held-out scene can still exhibit systematic offsets from material mismatch, unmodeled clutter, or simulator–reality gaps. When a few labeled links are available in that target scene, we therefore keep the pretrained encoder and task head frozen and attach a lightweight residual head that corrects only the residual error. Let ŷ0 denote the frozen prediction at a query. The adapted output is ŷ = ŷ0 + r(g),
(9)
where r is a small MLP and g collects TX–RX geometry used by the residual, including the RX location, the displacement q − xtx , and the link distance. For PG, r predicts a scalar correction in dB; for PAS, the same geometric features feed a compact spectrum head that outputs an additive angular residual map. The final layer of r is zero-initialized, so adaptation begins as an identity mapping and cannot degrade the frozen prior before training. Only the residual parameters are updated on the target-scene labels, which keeps adaptation cheap, preserves the cross-scene representation learned by the encoder, and applies unchanged to both PG and PAS. Training objective. Learning proceeds in stages that mirror this encoder–decoder factorization. We first pretrain the shared encoder and PG head with mean squared error on normalized PG targets, so the backbone absorbs a cross-scene propagation prior. For PAS, we then freeze the encoder and optimize only the directional head on spectrum labels. When target-scene labels are available, a final optional stage freezes both the encoder and the task head and fits only the residual r. This staging reuses one scene representation across tasks while confining expensive updates to the earliest pretraining stage.
4
Experiments
4.1
Datasets and Protocols
We construct PRISM to provide paired geometry and radio labels for cross-scene learning. Using Infinigen Raistrick et al. (2024), we generate 391 indoor rooms with mesh geometry, surface materials, and textures 6
(longest horizontal extents 3–16 m). Surface names map to 15 material classes with electromagnetic parameters from ITU-R P.2040 itu (2023) (Figure 3). Scenes are exported to XML and labeled in Sionna RT Hoydis et al. (2023) for PG and PAS. PRISM-PG. For PG, Sionna’s radio-map solver produces a dense 163 volume for each TX at 3.5 GHz, a representative mid-band 5G carrier, with up to 256 TX locations per room. After filtering invalid simulations we retain 337 rooms (269 / 33 / 35 train / val / test), totaling 86,272 TX-conditioned fields. Point clouds sampled from surfaces serve as model inputs; meshes are used only for offline labeling. PRISM-PAS. For PAS, we sample 20 × 20 TX–RX pairs per room, compute channel responses with Sionna’s path solver, and form power spectra by conventional beamforming on a 4 × 4 RX array (360 × 90 azimuth– elevation). Intersection with PRISM-PG yields 199 rooms that share mesh geometry but use independently sampled TX–RX grids. A denser held-out indoor room provides 2000 TX locations and 20 RXs for single-scene comparison.
4.2
Implementation and Baselines
Model and training. The reported PG checkpoint uses width d=256, CSC depth N =4, 16 heads, FPS M =20,000, downsampling rate 3, decoder neighborhood K=8, log10 σ material channels, and TX–query geometry (≈29.8 M parameters). PAS reuses the same backbone with a frozen encoder and Fourier direction encoding of L=6 frequencies. We optimize with Adam, cosine learning-rate schedule, initial learning rate 2 × 10−4 , MSE loss on normalized PG targets, batch size 16 with data parallelism, and up to 3000 epochs; the reported checkpoint is selected by best validation MAE. Scene splits and PAS train/test TX draws use seed 2026. During development we mainly compared widths {128, 256}, CSC depths {2, 4}, and neighborhood sizes K ∈ {8, 16}, and kept the setting with the best validation MAE. Unless noted, training and evaluation use PyTorch 2.4 on NVIDIA RTX 3090 GPUs; latency and peak-memory numbers in Table 4 are measured on a single RTX 3090 (batch size 1, warm steady state). Baselines. We compare against learning methods rather than scoring a ray tracer on its own labels. For cross-scene PG, we use two feedforward baselines trained on the same scene-disjoint split: a RadioUNet-style slice-wise 2D U-Net Levie et al. (2021) and a 3D U-Net on occupancy/material volumes. For per-scene comparisons, we train NeRF2 Zhao et al. (2023), WRF-GS+ Wen et al. (2026a), and GSRF Yang et al. (2026) on labeled links inside a target room and evaluate held-out TXs. NeRF2 and WRF-GS+ appear in both the PG sample-efficiency study and the PAS transfer study; GSRF is used for PAS. Metrics. We evaluate PG in dB after clipping predictions ŷ and labels y to [−90, 0] dB. Over N evaluated samples, MAE N 1 X |ŷi − yi | (10) MAE = N i=1 is our primary metric, since radio coverage is conventionally reported on a logarithmic power scale. We also report normalized mean squared error (NMSE) and PSNR, PSNR = 10 log10 1 PN N
R2
i=1 (ŷi − yi )
2
,
(11)
with dynamic range R=90 dB matching the clipped span. NMSE removes absolute-scale sensitivity, while PSNR converts the same error into a familiar logarithmic score. SSIM is computed on horizontal 16 × 16 slices of each 163 volume with the same range, to capture spatial structure beyond pointwise error. For PAS, predictions and labels are scored in a per-pair normalized spectrum space on [0, 1]. We use SSIM under unit dynamic range for angular-lobe structure and MAE for intensity error on the same maps. Tables report means over RXs, and CDFs provide median and percentile behavior beyond a single average.
7
Method
MAE↓
NMSE
PSNR↑
SSIM↑
RadioUNet (2D) 3D U-Net Point2Radio (ZS)
3.745 3.038
0.0159 0.0152
25.10 26.71
0.836 0.884
0.871
0.0025
34.83
0.954
Table 1 ZS full-volume PG prediction on the 35 held-out test scenes
1.0 101
SSIM ↑
MAE (dB) ↓
0.9 0.8 0.7
100
Point2Radio (FT) Point2Radio (ZS) WRF-GS+ NeRF²
0.6 1
10
100
200
1
Ntx
10
100
200
Ntx
Figure 4 PG prediction accuracy on three held-out rooms under different numbers of training TXs.
4.3
PG Prediction
Cross-scene feedforward baselines. We first compare against methods that, like Point2Radio, aim to generalize to unseen scenes. Table 1 reports full-volume PG on the 35 held-out test rooms. Point2Radio reaches 0.871 dB MAE and 0.954 SSIM (NMSE 0.0025, PSNR 34.83 dB), reducing MAE by 76.7% relative to RadioUNet (3.745 dB / 0.836) and remaining substantially more accurate than 3D U-Net (3.038 dB / 0.884). Pure 2D slice models cannot resolve the underlying 3D geometry, so even with height labels they struggle to capture vertical structure and multipath. A volumetric 3D U-Net provides a 3D representation, yet still learns a largely scene-agnostic input–output map rather than propagation structure, and therefore transfers poorly across rooms. Figure 1 visualizes the same gap on three held-out rooms. Per-scene neural fields and sample efficiency. We next compare against per-scene methods that do not transfer across rooms. NeRF2 and WRF-GS+ overfit a single environment from dense in-scene supervision and then reconstruct radio fields for arbitrary TXs in that room. Against this paradigm we report two operating modes for Point2Radio: ZS cross-scene inference, and light target-scene adaptation with a frozen backbone. To measure sample efficiency, we evaluate on three held-out rooms with 256 TXs each: a fixed set of 56 TXs is reserved for testing, and from the remaining 200 TXs we draw nested training sets of size Ntx ∈ {1, 10, 100, 200} for per-scene training (Figure 4). Point2Radio already provides strong ZS PG (≈1.15 dB MAE / 0.947 SSIM). As Ntx grows, NeRF2 and WRF-GS+ improve steadily; at Ntx =200, WRF-GS+ exceeds our ZS SSIM (0.953 vs. 0.947) while still trailing in MAE (1.57 dB vs. 1.15 dB). Fitting a lightweight residual head on the same training TXs further improves Point2Radio, reducing MAE to 0.83 dB and raising SSIM to 0.958 at Ntx =200 (∆MAE ≈0.32 dB, ∆SSIM ≈0.012 over ZS). Input ablation. In practice it is difficult to obtain point clouds that jointly provide accurate normals and electromagnetic material labels; materials are especially hard to infer, even with vision priors and light calibration An et al. (2026). Table 2 therefore removes ITU material channels and/or normals at test time on the same cross-scene PG split, keeping the trained weights fixed. Dropping materials raises MAE from
8
Variant
MAE (dB)↓
SSIM↑
0.871
0.954
1.259 1.456 1.685
0.943 0.933 0.920
Method
SSIM↑
MAE↓
NeRF2 WRF-GS+ GSRF Point2Radio (ZS) Point2Radio (residual FT)
0.6775 0.7145 0.5686 0.6899
0.0993 0.0872 0.1339 0.1128
Full Model w/o ITU materials w/o normals w/o materials & normals Table 2 Ablation of test-time inputs on cross-scene PG.
0.7983
0.0798
Table 3 PAS prediction on one held-out indoor room (4 RXs; per-RX 80/20 train/test TX split). Mean over RXs.
0.871 dB to 1.259 dB, dropping normals to 1.456 dB, and removing both to 1.685 dB; SSIM falls from 0.954 to 0.943 / 0.933 / 0.920. Normals thus matter more than material channels, consistent with their role in reflection and occlusion, while materials still help for strong reflectors such as metal. Even without either cue, Point2Radio remains well below the U-Net baselines in Table 1, showing that positions alone carry a usable geometric prior and that normals and ITU attributes act as refinements rather than hard prerequisites.
4.4
PAS Prediction
In PAS baselines, the RX is typically fixed while the TX moves, and models are fit with dense in-scene supervision to reconstruct the spectrum at that RX. We follow the same protocol in one held-out indoor room that is unseen during Point2Radio pretraining. The room provides four RXs and 2000 TX locations per RX; for each RX we use an 80/20 TX split (1600 train / 400 test) and report the mean over the four RXs. Per-scene NeRF2 , WRF-GS+, and GSRF are trained on the training TXs of each RX. For Point2Radio we evaluate two modes with a frozen PG-pretrained encoder: ZS inference, and light residual-head fine-tuning on the training TXs. Table 3 reports mean test SSIM and MAE over the four RXs. Figure 6 shows qualitative examples, and Figure 5 reports per-pair CDFs over the 1600 test pairs. ZS Point2Radio reaches mean SSIM 0.6899 (median 0.6925), slightly above NeRF2 (0.6775 / 0.6870) and well above GSRF (0.5686 / 0.5821), while remaining close to WRF-GS+ (0.7145 / 0.7259). After residual finetuning, Point2Radio rises to mean SSIM 0.7983 (median 0.8105, 90th 0.8866) and mean MAE 0.0798 (median 0.0745), exceeding WRF-GS+ by +0.084 mean SSIM and improving the 90th-percentile SSIM from 0.8206 to 0.8866. Relative to NeRF2 and GSRF the mean-SSIM gains are +17.8% and +40.4%. The spatial backbone stays frozen, so the gain comes from reusing the PG prior with a lightweight scene-specific correction. Together, the PG and PAS results support a consistent picture of what the shared encoder has learned. Cross-scene PG accuracy shows that the frozen backbone already captures how a TX interacts with 3D geometry and materials, rather than memorizing a single room. Reusing that encoder for PAS with only a lightweight task head then yields competitive ZS spectra and strong gains after residual adaptation, indicating that the same TX–scene interaction prior transfers across radio quantities. Overall, Point2Radio stays compact in parameters and inference cost, yet delivers high accuracy across tasks and scenes through a reusable propagation representation plus light task- and scene-specific heads.
4.5
Inference Efficiency
Table 4 compares feedforward Point2Radio with the Sionna RT labeling pipelines on an RTX 3090, batch size 1 in warm steady state. For one TX-conditioned 163 PG volume, Point2Radio finishes in 163 ms at 3.3 GB 9
CDF
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
0.5
1.0
SSIM
Point2Radio (FT) Point2Radio (ZS) WRF-GS+ NeRF² GSRF
0.0 0.0
0.1
MAE
0.2
0.3
Figure 5 CDF of per-pair test SSIM / MAE on one held-out indoor room (4 RXs).
PG (163 )
PAS (1 pair)
Method
Time
Peak GPU
Time
Peak GPU
Our encode Our decode Our end-to-end Sionna RT
151 ms 10 ms 163 ms 990 ms
— — 3.3 GB 8.3 GB
150 ms 3 ms 156 ms 648 ms
— — 1.2 GB 3.3 GB
Table 4 Inference latency and peak GPU memory on an RTX 3090.
peak memory, versus 990 ms / 8.3 GB for Sionna, about 6.1× faster and 2.5× less memory. For one PAS pair the gap is 156 ms / 1.2 GB versus 648 ms / 3.3 GB, about 4.2× faster. Inference uses only a point cloud and transceiver queries, with no mesh and no online path tracing. The speedup comes from where the compute sits. Encoding a TX-conditioned scene representation takes about 151 ms, while decoding a full 163 volume from the cached codes takes only about 10 ms, and a single RX query about 3 ms. Because most of the work is in the encoder and the local decoder is cheap, dense querying under a fixed TX does not grow linearly with the number of RXs. The model encodes once and then evaluates many locations quickly on the decoder side. Combined with the PG accuracy above, this makes Point2Radio attractive as a fast surrogate for large-scale radio-field generation. Regenerating the 86,272 TX-conditioned volumes in PRISM-PG at these steady-state rates would take about 24 hours with Sionna RT, but only about 4 hours with Point2Radio, a roughly 6× reduction in wall-clock labeling time while retaining high fidelity on held-out scenes.
5
Conclusion and Future Work
We present Point2Radio, a foundation model that learns transferable TX-conditioned radio fields from materialaware point clouds. Hierarchical tokenization and CSC attention produce a reusable scene representation, which task-specific query heads map to dense PG fields and PAS, with light residual adaptation when target labels are available. This suggests that, with sufficiently diverse data, a model can capture accurate propagation structure without explicitly executing deterministic physics at inference. Our experiments show strong cross-scene accuracy and efficiency. Point2Radio achieves 0.871 dB MAE and 0.954 SSIM for PG, reducing error by 76.7% relative to a UNet-style baseline, transfers effectively to PAS, and runs several times faster than Sionna RT at lower peak memory, without meshes or online path tracing. Future work will focus on the following directions. 10
GT | SSIM | MAE
NeRF² | 0.587 | 0.097
0°
315°90
WRF-GS+ | 0.595 | 0.120
0°
45°
315°90
60
0°
45°
315°90
60 30
45° 60
30
270°
90°
225°
270°
135°
90°
225°
270°
135°
180°
90°
225° 180°
Point2Radio (ZS) | 0.660 | 0.126 Point2Radio (FT) | 0.860 | 0.052
0°
0°
45°
315°90
60
0°
45°
315°90
60 30
90°
135° 180°
45° 60
30
270°
225°
135°
180°
GSRF | 0.577 | 0.110 315°90
30
30
270°
90°
225°
135° 180°
270°
90°
225°
135° 180°
Figure 6 PAS qualitative examples on held-out TX–RX pairs in one indoor room.
• Input modality. Our pipeline currently assumes a prepared point cloud, which is easier than a simulation mesh but still nontrivial in practical deployments. Future work may take more accessible modalities such as images and videos as input, enabling Point2Radio to support a broader range of downstream tasks. • Training and adaptation data. We aim to train on larger, higher-quality multimodal corpora to reduce dependence on simulator fidelity, and to use real radio measurements for rapid fine-tuning across diverse downstream tasks.
References Effects of building materials and structures on radiowave propagation above about 100 mhz. Technical Report Recommendation ITU-R P.2040-3, International Telecommunication Union, August 2023. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. Zhenlin An, Longfei Shangguan, John Kaewell, Philip Pietraski, and Kyle Jamieson. Radiotwin: A digital building material twin for wideband, cross-link, cross-band wireless channel prediction. In 2025 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), pages 1–10, 2025. doi: 10.1109/DySPAN64764.2025.11115919. Zhenlin An, Longfei Shangguan, John Kaewell, Philip Pietraski, Jelena Senic, Camillo Gentile, Nada Golmie, and Kyle Jamieson. Taming vision priors for data efficient mmwave channel modeling. arXiv preprint arXiv:2603.13383, 2026. Ge Cao and Zhen Peng. Raypronet: A neural point field framework for radio propagation modeling in 3d environments. IEEE Journal on Multiscale and Multiphysics Computational Techniques, 9:330–340, 2024. doi: 10.1109/JMMCT. 2024.3464373.
11
Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts. arXiv preprint arXiv:2511.16719, 2025. Xingyu Chen, Zihao Feng, Ke Sun, Kun Qian, and Xinyu Zhang. Rfcanvas: Modeling rf channel by fusing visual priors and few-shot rf measurements. In Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems, SenSys ’24, page 464–477, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 9798400706974. doi: 10.1145/3666025.3699351. URL https://doi.org/10.1145/3666025.3699351. Xingyu Chen, Zihao Feng, Kun Qian, and Xinyu Zhang. Radio frequency ray tracing with neural object representation for enhanced rf modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21339–21348, June 2025. Heyu Guo, Shanmu Wang, Ruichun Ma, Shiqi Jiang, Yasaman Ghasempour, Omid Abari, Baining Guo, and Lili Qiu. Omnivla: Physically-grounded multimodal vla with unified multi-sensor perception for robotic manipulation. arXiv preprint arXiv:2511.01210, 2025. Jakob Hoydis, Fayçal Aït Aoudia, Sebastian Cammerer, Merlin Nimier-David, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. Sionna RT: Differentiable ray tracing for radio propagation modeling. In IEEE Globecom Workshops (GC Wkshps), pages 317–321, 2023. Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. on Graphics, 42(4), Jul. 2023. Ju-Hyung Lee, Omer Gokalp Serbetci, Dheeraj Panneer Selvam, and Andreas F. Molisch. Pmnet: Robust pathloss map prediction via supervised learning. In GLOBECOM 2023 - 2023 IEEE Global Communications Conference, pages 4601–4606, 2023. doi: 10.1109/GLOBECOM54140.2023.10437562. Ron Levie, Cagkan Yapar, Gitta Kutyniok, and Giuseppe Caire. Radiounet: Fast radio map estimation with convolutional neural networks. IEEE Transactions on Wireless Communications, 20(6):4001–4015, 2021. doi: 10.1109/TWC.2021.3054977. Boxun Liu, Shijian Gao, Xuanyu Liu, Xiang Cheng, and Liuqing Yang. Wifo: Wireless foundation model for channel prediction. Science China Information Sciences, 68(6):162302, 2025. Haofan Lu, Christopher Vattheuer, Baharan Mirzasoleiman, and Omid Abari. Newrf: a deep learning framework for wireless radiation field reconstruction and channel prediction. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024. Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021. Tribhuvanesh Orekondy, Pratik Kumar, Shreya Kadambi, Hao Ye, Joseph Soriaga, and Arash Behboodi. WiNeRT: Towards neural ray tracing for wireless channel modelling and differentiable simulations. In The Eleventh International Conference on Learning Representations (ICLR), 2023. URL https://openreview.net/forum?id=tPKKXeW33YU. Alexander Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, Zeyu Ma, and Jia Deng. Infinigen indoors: Photorealistic indoor scenes using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21783–21794, June 2024. Tapan K Sarkar, Zhong Ji, Kyungjung Kim, Abdellatif Medouri, and Magdalena Salazar-Palma. A survey of various propagation models for mobile communication. IEEE Antennas and Propagation Magazine, 45(3):51–82, Mar. 2003. Shen Wang, Guosheng Wang, Junyang Liu, Donghui Dai, and Lei Yang. Sign-rf: Self-adaptive neural fields for scalable urban radio reconstruction. In IEEE INFOCOM 2026 - IEEE Conference on Computer Communications, pages 1–10, 2026a. doi: 10.1109/INFOCOM59046.2026.11571284. Sheng Wang, Hengtao He, Chaozheng Wen, Jingwen Tong, Xinyu Li, Xiao Li, Jun Zhang, and Shi Jin. Xfreq-gs: Cross-frequency wireless radiation field reconstruction with 3d gaussian splatting. arXiv e-prints, pages arXiv–2605, 2026b. Xiucheng Wang, Qiming Zhang, Nan Cheng, Junting Chen, Zezhong Zhang, Zan Li, Shuguang Cui, and Xuemin Shen. Radiodiff-3d: A 3d× 3d radio map dataset and generative diffusion based benchmark for 6g environmentaware communication. IEEE Transactions on Network Science and Engineering, 13:3773–3789, 2026c. doi: 10.1109/TNSE.2025.3590545.
12
Chaozheng Wen, Jingwen Tong, Yingdong Hu, Zehong Lin, and Jun Zhang. WRF-GS: Wireless radiation field reconstruction with 3D Gaussian splatting. In IEEE Conference on Computer Communications (INFOCOM), pages 1–10, 2025. Chaozheng Wen, Jingwen Tong, Yingdong Hu, Zehong Lin, and Jun Zhang. Neural representation for wireless radiation field reconstruction: A 3d gaussian splatting approach. IEEE Transactions on Wireless Communications, 25: 7490–7504, 2026a. doi: 10.1109/TWC.2025.3631663. Chaozheng Wen, Jingwen Tong, Zehong Lin, Chenghong Bian, and Jun Zhang. Bridging visual and wireless sensing: A unified radiation field for 3d radio map construction. arXiv preprint arXiv:2601.19216, 2026b. Bing Xu, Mukund Varma T, Cheng Wang, Tzu-mao Li, Lifan Wu, Bartlomiej Wronski, Ravi Ramamoorthi, and Marco Salvi. A generalizable light transport 3d embedding for global illumination. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, SIGGRAPH Conference Papers ’26, New York, NY, USA, 2026. Association for Computing Machinery. ISBN 9798400725548. doi: 10.1145/3799902.3811095. URL https://doi.org/10.1145/3799902.3811095. Kang Yang, Gaofeng Dong, Sijie Ji, Wan Du, and Mani Srivastava. Gsrf: Complex-valued 3d gaussian splatting for efficient radio-frequency data synthesis. Advances in Neural Information Processing Systems, 38:63076–63105, 2026. Yong Zeng, Junting Chen, Jie Xu, Di Wu, Xiaoli Xu, Shi Jin, Xiqi Gao, David Gesbert, Shuguang Cui, and Rui Zhang. A tutorial on environment-aware communications via channel knowledge map for 6g. IEEE Communications Surveys & Tutorials, 26(3):1478–1519, 2024. doi: 10.1109/COMST.2024.3364508. Lihao Zhang, Haijian Sun, Samuel Berweger, Camillo Gentile, and Rose Qingyang Hu. Rf-3dgs: Wireless channel modeling with radio radiance field and 3d gaussian splatting. IEEE Transactions on Wireless Communications, 25: 10419–10433, 2026a. doi: 10.1109/TWC.2026.3652154. Yumeng Zhang, Jiajia Guo, Chaozheng Wen, Chenghong Bian, and Jun Zhang. Geogs-ce: Learning delay–beam channel priors with 3d gaussians for high-mobility scenarios. arXiv preprint arXiv:2605.16094, 2026b. Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. NeRF2: Neural radio-frequency radiance fields. In International Conference on Mobile Computing and Networking (MOBICOM), pages 1–15, 2023. Tianyue Zheng, Jiajia Guo, Linglong Dai, Shi Jin, and Jun Zhang. Muse-fm: Multi-task environment-aware foundation model for wireless communications. IEEE Transactions on Wireless Communications, 2026.
13