LAETwin-XL: Low-Altitude XL-MIMO Framework

• LAETwin-XL is a digital twin-based XL-MIMO framework for low-altitude aerial communications that integrates realistic 3D environment reconstruction, ray-tracing channel synthesis, and a comprehensive dataset.
• It employs a conditional denoising diffusion model with a Transformer backbone to reconstruct full hybrid-field channels from sparse antenna observations, achieving superior extrapolation, classification, and localization performance.
• The framework addresses challenges in near-field propagation, 3D mobility, and limited high-fidelity measurements by leveraging urban and suburban scenarios for pretraining and fine-tuning across multiple tasks.

Searching arXiv for the specified paper and a few directly related references so the article can cite them accurately. arXiv search query: (Li et al., 12 Jun 2026) LAETwin-XL denotes a low-altitude XL-MIMO research framework that combines a digital twin-based channel generation toolchain and dataset with a conditional denoising diffusion implicit model (CDDIM)-based generative foundation model for incomplete channel observations in low-altitude economy scenarios (Li et al., 12 Jun 2026). In the formulation presented in "Digital Twin-Based Channel Generation Toolchain and Foundation Model for Low-Altitude XL-MIMO" (Li et al., 12 Jun 2026), the framework is designed to address the joint difficulty created by three-dimensional mobility, near-field propagation, and the lack of high-fidelity wireless datasets for low-altitude aerial communication systems. Its scope therefore spans physically grounded environment reconstruction, ray-tracing-based hybrid-field channel synthesis, pretraining on sparse antenna observations, zero-shot channel extrapolation, and parameter-efficient transfer to downstream tasks including channel estimation, classification, and localization (Li et al., 12 Jun 2026).

1. Conceptual scope and motivation

LAETwin-XL is defined as a complete framework for low-altitude XL-MIMO research with three tightly coupled components: a digital twin toolchain that constructs realistic 3D city and suburban environments from OpenStreetMap, simulates UAV trajectories, and uses Sionna’s differentiable ray-tracing engine to generate per-antenna near-field and far-field channels for a $64\times 64$ UPA at a $7\ \text{GHz}$ carrier; the LAETwin-XL dataset generated by that toolchain; and a CDDIM-based generative foundation model pretrained to reconstruct full XL-MIMO channels from sparse antenna observations and adapted to several downstream tasks (Li et al., 12 Jun 2026).

The motivation arises from low-altitude economy applications such as UAVs, eVTOLs, and drones for logistics and surveillance, which require robust air-ground links under full 3D mobility and near-field XL-MIMO operation (Li et al., 12 Jun 2026). The paper identifies four central challenges. First, near-field propagation induces spherical wavefronts, element-dependent path distances and gains, and spatial non-stationarity, which are not captured by standard plane-wave, far-field datasets such as DeepMIMO (Li et al., 12 Jun 2026). Second, aerial users move through realistic 3D geometry with varying altitude and trajectory structure, including road-following motion in urban environments and almost-linear cruising in suburban areas (Li et al., 12 Jun 2026). Third, the authors argue that existing open-source measured or simulated datasets, including DeepSense 6G, DeepMIMO, CAVIAR6G, DeepVerse 6G, and BUPTCMCC-6G, do not jointly provide XL-MIMO-scale arrays, 3D low-altitude mobility, near-field modeling, and digital-twin realism (Li et al., 12 Jun 2026). Fourth, existing wireless foundation models such as LWM and LWLM are described as primarily targeting terrestrial massive MIMO, employing masked autoencoding with moderate mask ratios, and relying on relatively complete input channels, whereas practical LAE XL-MIMO may operate with sparse antenna activation to reduce RF chains and power (Li et al., 12 Jun 2026).

A plausible implication is that LAETwin-XL is positioned not merely as a dataset release but as an attempt to standardize a physically grounded pretraining regime for hybrid-field aerial channels. The paper explicitly frames the framework as filling two gaps at once: a standardized digital twin-based LAE dataset that captures hybrid-field channels in realistic 3D city and suburban environments, and a generative foundation model specialized for XL-MIMO channel extrapolation and downstream tasks from incomplete observations (Li et al., 12 Jun 2026).

2. Low-altitude XL-MIMO system model

The physical scenario is an uplink LAE system in which a ground base station with center at $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$ is equipped with a $64\times 64$ UPA in the $y$-$z$ plane, facing $+x$, while the user equipment is a single-antenna UAV transmitter at position $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$ (Li et al., 12 Jun 2026). The base-station height is $(0,0,35\ \text{m})$ in urban scenarios and $(0,0,65\ \text{m})$ in suburban scenarios. UAV altitude is uniformly sampled in $7\ \text{GHz}$0 for urban environments and in $7\ \text{GHz}$1 for suburban environments (Li et al., 12 Jun 2026). The communication setting uses center frequency $7\ \text{GHz}$2, an OFDM system with subcarrier spacing $7\ \text{GHz}$3, and, for most experiments, the channel at the carrier frequency and the first OFDM symbol of each snapshot (Li et al., 12 Jun 2026).

The array geometry is specified through the antenna positions

$7\ \text{GHz}$4

where $7\ \text{GHz}$5, $7\ \text{GHz}$6, and $7\ \text{GHz}$7 are inter-element spacings consistent with 3GPP TR 38.901 (Li et al., 12 Jun 2026). Near-field behavior is characterized using the Rayleigh distance

$7\ \text{GHz}$8

with $7\ \text{GHz}$9 the UPA aperture. For a $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$0 UPA at $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$1, the paper states that $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$2 is on the order of tens of meters, so UAVs at $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$3–$\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$4 can often lie inside the near-field region and may transition across near-field and far-field zones along a single trajectory (Li et al., 12 Jun 2026).

Mobility is modeled separately for urban and suburban regimes. Urban operations assume UAVs follow roads for use cases such as vehicle tracking and road inspection, with horizontal trajectories generated by SUMO from road networks derived from OSM and controlled by parameters such as period and maxSpeed (Li et al., 12 Jun 2026). Their altitude evolves by bounded random walk or sinusoidal variation after a start altitude is drawn from $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$5. Suburban cruising uses near-linear horizontal motion with random perturbations to an initial velocity, combined with the same random-walk or sinusoidal altitude modes (Li et al., 12 Jun 2026). The maximum UAV speed is $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$6, yielding coherence time $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$7, while the snapshot interval is $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$8, so snapshots are treated as effectively independent in time (Li et al., 12 Jun 2026).

This system model is significant because the paper does not abstract away the field regime. Instead, the channel synthesis is structured to preserve the fact that a meter-scale array at mid-band frequency can operate in a hybrid-field regime for aerial users in realistic low-altitude trajectories (Li et al., 12 Jun 2026).

3. Digital twin toolchain and channel synthesis

The digital twin toolchain is organized into three stages: static environment construction, dynamic-scene generation with UAV trajectories, and Sionna RT-based channel simulation (Li et al., 12 Jun 2026). In the first stage, a region of interest is selected via OSM, with roads, building footprints, and heights extracted from the map source; OSM coordinates in WGS84 are converted to a local planar coordinate system such as Gauss–Kruger and then translated so that the ROI centroid becomes the origin, producing a metric 3D coordinate frame for each scenario (Li et al., 12 Jun 2026). Buildings are extruded using heights from OSM attributes or user-defined defaults, and surfaces are assigned ITU-compliant materials supported by Sionna RT, including itu-concrete, itu-metal, itu-medium-dry-ground, and itu-glass (Li et al., 12 Jun 2026). These steps define a 3D static environment $\mathbf{o}_{\mathrm R}\in\mathbb{R}^3$9 with geometry and electromagnetic properties.

The second stage inserts the SUMO-based 3D trajectories into $64\times 64$0 to produce a dynamic scene $64\times 64$1, with explicit coordinate transformations to ensure co-registration between UAV positions and building geometry (Li et al., 12 Jun 2026). Trajectories are generated in SUMO’s local frame, transformed back to WGS84, and then mapped into the ROI-centered Cartesian coordinate frame shared by the static environment (Li et al., 12 Jun 2026).

The third stage uses Sionna RT to simulate propagation from discrete UAV positions $64\times 64$2 to every array element (Li et al., 12 Jun 2026). For each antenna element $64\times 64$3, the ray tracer computes propagation paths using shooting and bouncing rays, the image method, and hashing-based deduplication, returning a set

$64\times 64$4

where $64\times 64$5 is the number of paths to antenna $64\times 64$6, $64\times 64$7 is total path length, $64\times 64$8 is complex path gain, and $64\times 64$9 includes azimuth and elevation AoA/AoD at each interaction (Li et al., 12 Jun 2026). For a line-of-sight path,

$y$0

while for an NLoS path with $y$1 scatterers $y$2,

$y$3

(Li et al., 12 Jun 2026). Doppler is modeled using the UAV velocity $y$4 and path departure direction $y$5,

$y$6

and the complex path gain is written as

$y$7

with polarization field patterns $y$8 and interaction transfer matrices $y$9 (Li et al., 12 Jun 2026).

The resulting element-wise near-field channel is modeled as

$z$0

so each antenna experiences its own path lengths, gains, and Dopplers (Li et al., 12 Jun 2026). In the far-field regime, the paper states that the channel reduces to a plane-wave model with antenna-independent path gains and incident directions, making far-field behavior a limiting case of the same ray-tracing procedure (Li et al., 12 Jun 2026). The toolchain outputs full per-antenna channel vectors $z$1 or reshaped grids $z$2, together with path-level labels, Doppler, UE positions and velocities, environment indices and material metadata, and near-/far-field labels based on Rayleigh distance (Li et al., 12 Jun 2026).

A plausible implication is that the toolchain treats hybrid-field behavior as an emergent property of scene geometry and array scale rather than as a post hoc label. That design choice is central to the dataset’s use for tasks such as extrapolation and localization, where physically consistent depth and angular structure matter (Li et al., 12 Jun 2026).

4. Dataset construction and labeling

The LAETwin-XL dataset is divided into a large pretraining set and a smaller downstream set (Li et al., 12 Jun 2026). The pretraining dataset contains train, validation, and test cities and totals $z$3 samples (Li et al., 12 Jun 2026). The training partition includes urban scenes from Shanghai, Tianjin, Hangzhou, Shenzhen, New York, London, and Tokyo, and a suburban scene from Sydney; validation uses Singapore for both urban and suburban settings; test uses Dubai for urban and suburban settings (Li et al., 12 Jun 2026). The paper gives per-city counts, including Shanghai with 10 urban scenes and $z$4 samples, of which $z$5 are near-field and $z$6 are far-field (Li et al., 12 Jun 2026). Each scene includes multiple UAVs and 10 time-instants per trajectory (Li et al., 12 Jun 2026).

The downstream dataset is intended for fine-tuning and evaluation. It uses Beijing for training, Berlin for validation, and Nanjing for test, each with urban and suburban scenes, and totals $z$7 samples across splits (Li et al., 12 Jun 2026). Each sample corresponds to one UAV snapshot and one base-station UPA channel snapshot for effectively one OFDM symbol at $z$8 (Li et al., 12 Jun 2026). The dataset design therefore explicitly tests cross-city generalization by separating pretraining cities from downstream cities (Li et al., 12 Jun 2026).

The provided labels are unusually rich for an XL-MIMO dataset. For each sample, the dataset includes the complex XL-MIMO channel vector $z$9, reshaped as $+x$0; UE metadata such as 3D position $+x$1 and possibly velocity $+x$2; path-level labels from Sionna RT, including $+x$3, $+x$4, AoA/AoD at interactions, and the number of interactions $+x$5; and environment labels such as city index, scene ID, urban/suburban tag, and near-field versus far-field label for the UE–BS relation (Li et al., 12 Jun 2026).

Component	Content	Purpose
Pretraining dataset	$+x$6 samples across 13 cities and settings	Foundation-model pretraining
Downstream dataset	$+x$7 samples across Beijing, Berlin, and Nanjing splits	Fine-tuning and evaluation
Per-sample labels	Channel, geometry, path-level parameters, field-regime labels	Multi-task supervision

This labeling supports channel extrapolation, channel estimation, near-/far-field classification, and LoS-based localization directly, while the paper also notes that the same structure could support other tasks such as beam prediction and sensing (Li et al., 12 Jun 2026). This suggests that LAETwin-XL is organized less as a single-benchmark corpus than as a supervised and self-supervised substrate for multiple aerial XL-MIMO learning problems.

5. CDDIM foundation model and learning from partial observations

The foundation model operates over the antenna-domain channel matrix $+x$8 and is conditioned on partial observations defined by a binary mask $+x$9, where $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$0 marks an observed antenna and 0 marks a masked antenna (Li et al., 12 Jun 2026). The masked clean channel is

$\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$1

the mask ratio $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$2 is the fraction of zeros, and the antenna selection ratio is $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$3 (Li et al., 12 Jun 2026). During pretraining, $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$4 is randomly sampled from $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$5, so the model is trained with conditions ranging from moderate to severe sparsity (Li et al., 12 Jun 2026).

Forward diffusion follows a variance-preserving process

$\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$6

where $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$7, $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$8, $\mathbf{o}_{\mathrm T}\in\mathbb{R}^3$9, and $(0,0,35\ \text{m})$0 is a linear schedule from $(0,0,35\ \text{m})$1 to $(0,0,35\ \text{m})$2 over $(0,0,35\ \text{m})$3 steps (Li et al., 12 Jun 2026). The denoiser $(0,0,35\ \text{m})$4 uses a Transformer backbone rather than a CNN, motivated in the paper by near-field non-stationarity and global correlations caused by spherical wavefront curvature (Li et al., 12 Jun 2026).

The denoiser takes as input the noisy full channel $(0,0,35\ \text{m})$5, the partial clean channel $(0,0,35\ \text{m})$6, the mask $(0,0,35\ \text{m})$7, and diffusion time $(0,0,35\ \text{m})$8, embedding them into patch tokens and positional encodings (Li et al., 12 Jun 2026). A masked multi-head cross-attention mechanism computes

$(0,0,35\ \text{m})$9

where $(0,0,65\ \text{m})$0 contributes 0 for active antennas and $(0,0,65\ \text{m})$1 for masked entries, thereby forcing attention to ignore unobserved zero-filled positions (Li et al., 12 Jun 2026). Geometry-aware positional encodings include a 2D sinusoidal encoding for UPA indices $(0,0,65\ \text{m})$2 and a diffusion time encoding produced by sinusoidal embedding plus MLP (Li et al., 12 Jun 2026). The training objective is the MSE noise-prediction loss

$(0,0,65\ \text{m})$3

(Li et al., 12 Jun 2026).

Reverse sampling is deterministic via CDDIM. Starting from Gaussian noise $(0,0,65\ \text{m})$4, the model iterates from $(0,0,65\ \text{m})$5 to 0 in steps of $(0,0,65\ \text{m})$6, with the paper using $(0,0,65\ \text{m})$7, i.e. 50 reverse steps (Li et al., 12 Jun 2026). The update is given by

$(0,0,65\ \text{m})$8

with

$(0,0,65\ \text{m})$9

yielding a sample $7\ \text{GHz}$00 of the full channel consistent with the partial observation (Li et al., 12 Jun 2026).

The paper contrasts this setup with BERT-style masked autoencoding and argues that the diffusion formulation supports high mask ratios, avoids over-smoothing, and learns the full channel distribution rather than a single deterministic average (Li et al., 12 Jun 2026). This suggests that the generative prior is intended to encode near-field curvature and multipath structure as latent regularities of the channel manifold rather than as merely interpolative spatial patterns.

6. Zero-shot extrapolation, downstream transfer, and empirical behavior

Zero-shot channel extrapolation is defined as applying the pretrained CDDIM model without fine-tuning to generate $7\ \text{GHz}$01 from a clean partial channel $7\ \text{GHz}$02 and mask $7\ \text{GHz}$03 on test samples from a city in the pretraining domain, specifically Dubai in the reported experiments (Li et al., 12 Jun 2026). The test split contains Dubai urban and suburban samples totaling 4,000 examples, with mask ratios $7\ \text{GHz}$04, and evaluation relies mainly on NMSE, with PSNR and SSIM used for visual analysis (Li et al., 12 Jun 2026). Baselines are LWM with the same Transformer backbone trained via masked channel modeling, CDDIM with a U-Net backbone, CGAN, and zero-filling (Li et al., 12 Jun 2026). The paper reports that the proposed model dominates across all mask ratios and reaches NMSE of approximately $7\ \text{GHz}$05 at $7\ \text{GHz}$06, while LWM can be up to about $7\ \text{GHz}$07 worse (Li et al., 12 Jun 2026). Visual analyses in the amplitude, angular, and distance domains indicate that the model better preserves global structure, beam shape, and near-field distance focusing than the baselines (Li et al., 12 Jun 2026).

For downstream transfer, the pretrained foundation model serves as a universal backbone to which lightweight task heads are attached (Li et al., 12 Jun 2026). The fine-tuning strategy freezes the first 7 of 10 Transformer blocks and updates only the top 3 blocks plus the task head using limited downstream data (Li et al., 12 Jun 2026). A multi-time feature fusion mechanism extracts feature maps from five diffusion steps $7\ \text{GHz}$08, concatenates them along the token dimension, processes them with a Transformer encoder layer, and globally averages them to obtain a global feature vector $7\ \text{GHz}$09 for classification or localization (Li et al., 12 Jun 2026). Ablation results show that using 5 time steps yields the best downstream classification accuracy of $7\ \text{GHz}$10, outperforming single-step and 3-step variants (Li et al., 12 Jun 2026).

In channel estimation, only $7\ \text{GHz}$11 antennas observe pilots, with pilot-domain model

$7\ \text{GHz}$12

and the partial pilot grid $7\ \text{GHz}$13 is passed to a U-Net CE head through

$7\ \text{GHz}$14

(Li et al., 12 Jun 2026). The CE head outputs a noise residual, producing

$7\ \text{GHz}$15

which is then fed into the diffusion model for extrapolation (Li et al., 12 Jun 2026). End-to-end training uses

$7\ \text{GHz}$16

with

$7\ \text{GHz}$17

to enforce denoising consistency on observed antennas (Li et al., 12 Jun 2026). Relative to LS, ChannelNet, LWM+CE, and a train-from-scratch+CE baseline, the proposed model achieves the lowest NMSE across all tested SNRs and antenna selection ratios, with especially strong robustness at low SNR and low $7\ \text{GHz}$18 (Li et al., 12 Jun 2026).

For near-/far-field classification, the task is binary, the classifier uses focal loss with focusing parameter $7\ \text{GHz}$19, and the reported accuracy remains $7\ \text{GHz}$20 at $7\ \text{GHz}$21 and $7\ \text{GHz}$22 at $7\ \text{GHz}$23, a drop of only $7\ \text{GHz}$24 (Li et al., 12 Jun 2026). LWM with a classifier head is reported at about $7\ \text{GHz}$25 for $7\ \text{GHz}$26 and about $7\ \text{GHz}$27 for $7\ \text{GHz}$28, while training from scratch declines from about $7\ \text{GHz}$29 to about $7\ \text{GHz}$30 across the same range (Li et al., 12 Jun 2026).

For near-field localization, the experiments use a LoS-dominant near-field suburban subset with 450 training, 443 validation, and 386 test samples, and seek to estimate LoS path distance $7\ \text{GHz}$31, azimuth $7\ \text{GHz}$32, and elevation $7\ \text{GHz}$33 (Li et al., 12 Jun 2026). The localization head is an MLP that outputs cosine-sine angle representations and a raw distance variable, with angle recovery via stable atan2-like mappings and distance mapped into $7\ \text{GHz}$34 by a sigmoid-like transform (Li et al., 12 Jun 2026). Baselines include LWM+LOC, training from scratch, and OMP with a polar-domain near-field codebook of size $7\ \text{GHz}$35 built from a $7\ \text{GHz}$36 grid over azimuth, elevation, and distance $7\ \text{GHz}$37, with complexity scaling as $7\ \text{GHz}$38 (Li et al., 12 Jun 2026). The proposed model reports elevation MAE no greater than $7\ \text{GHz}$39, azimuth MAE no greater than $7\ \text{GHz}$40, and distance MAE no greater than $7\ \text{GHz}$41 even at $7\ \text{GHz}$42, whereas OMP exhibits similar angular MAE but distance MAE of $7\ \text{GHz}$43–$7\ \text{GHz}$44 (Li et al., 12 Jun 2026).

Taken together, these results indicate that the pretrained representation is useful both generatively and discriminatively. This suggests that the model’s latent structure is not limited to channel completion but retains information relevant to field-regime identification and geometric inference under sparse observations.

7. Training configuration, limitations, and prospective extensions

The pretraining configuration uses a 10-block Transformer backbone with approximately $7\ \text{GHz}$45 million parameters, $7\ \text{GHz}$46 diffusion steps, a linear $7\ \text{GHz}$47 schedule from $7\ \text{GHz}$48 to $7\ \text{GHz}$49, CDDIM reverse sampling with $7\ \text{GHz}$50, 2000 training epochs, batch size 256, AdamW, learning rate $7\ \text{GHz}$51, and mask ratio $7\ \text{GHz}$52 uniformly sampled from $7\ \text{GHz}$53 in each iteration (Li et al., 12 Jun 2026). Downstream fine-tuning runs for 300 epochs, freezes the first $7\ \text{GHz}$54 Transformer blocks, unfreezes the last $7\ \text{GHz}$55, sets the learning rate of unfrozen backbone blocks to $7\ \text{GHz}$56, and sets task-head learning rates to $7\ \text{GHz}$57 (Li et al., 12 Jun 2026). The CE head is a U-Net with 4 down and 4 up stages, the multi-time fusion module has about $7\ \text{GHz}$58 million parameters, and experiments are conducted on dual RTX 4090 GPUs (Li et al., 12 Jun 2026).

Ablation results characterize the principal performance–latency trade-offs. For reverse steps, $7\ \text{GHz}$59 corresponding to 200 steps achieves NMSE of about $7\ \text{GHz}$60 but with latency of about $7\ \text{GHz}$61, whereas $7\ \text{GHz}$62 corresponding to 50 steps gives NMSE of about $7\ \text{GHz}$63 and latency of about $7\ \text{GHz}$64, which the authors choose as a compromise (Li et al., 12 Jun 2026). Increasing backbone depth from $7\ \text{GHz}$65 to $7\ \text{GHz}$66 improves NMSE from $7\ \text{GHz}$67 to $7\ \text{GHz}$68, while $7\ \text{GHz}$69 reaches $7\ \text{GHz}$70; the 10-block model is selected as a balance between performance and model size (Li et al., 12 Jun 2026). For fine-tuning, $7\ \text{GHz}$71 yields the best classification accuracy at $7\ \text{GHz}$72, while $7\ \text{GHz}$73 shows reduced performance interpreted in the paper as excessive forgetting and $7\ \text{GHz}$74 as underfitting (Li et al., 12 Jun 2026).

The framework is also bounded by several explicit assumptions. The entire dataset is simulation-based and does not include measured channels; hardware impairments, calibration errors, and mutual coupling are not modeled; the dataset is restricted to the mid-band $7\ \text{GHz}$75 frequency; the experiments focus on the channel at the carrier frequency and neglect time variation within one OFDM symbol; temporal correlation across snapshots is not exploited because the snapshots are spaced by 1 s; and localization experiments are confined to LoS-dominant near-field suburban scenarios rather than heavily NLoS or blocked settings (Li et al., 12 Jun 2026). The paper further notes that ray tracing is computationally heavy and that diffusion sampling with 50 steps adds per-sample latency of roughly $7\ \text{GHz}$76 for extrapolation and channel estimation on the reported hardware (Li et al., 12 Jun 2026).

The stated future directions are real-world validation using XL-MIMO testbed measurements when available, and accelerated sampling through fewer steps, improved ODE solvers, or distillation to reduce latency (Li et al., 12 Jun 2026). The paper’s broader context also suggests possible extensions to other bands, richer propagation conditions, hardware impairments, and more complex scenarios such as multi-UAV, multi-cell, or joint communications and sensing, although these appear as contextual suggestions rather than reported results (Li et al., 12 Jun 2026).

The code and dataset are available at the repository specified by the paper, which includes scripts for 3D environment construction from OSM, SUMO trajectory integration, Sionna RT simulation, pretrained CDDIM weights, and pre-split pretraining and downstream datasets (Li et al., 12 Jun 2026). Within the architecture and experiments described, LAETwin-XL therefore functions simultaneously as a synthetic data-generation infrastructure, a benchmark substrate for low-altitude hybrid-field XL-MIMO, and a pretrained generative backbone for sparse-observation inference (Li et al., 12 Jun 2026).