XL-VLA: Scalable Vision-Language-Action & Communication

• XL-VLA is a scalable framework uniting vision-language-action models for robotics through a shared Transformer backbone with embodiment-specific soft prompts for cross-embodiment generalization.
• In wireless communications, XL-array designs utilize modular, high-resolution antenna arrays and non-uniform spherical wave models to achieve precise beam focusing and grating lobe suppression.
• Empirical evaluations show that techniques like flow-matching and prompt-based adaptation in XL-VLA significantly improve performance benchmarks across diverse robotic and communication scenarios.

XL-VLA (Extremely Large-Scale Vision-Language-Action Systems) refers to two distinct, high-impact domains in contemporary research: (1) scalable vision-language-action (VLA) models for robotics that leverage soft-prompted Transformers to achieve state-of-the-art cross-embodiment generalization, and (2) extremely large-scale antenna array (XL-array) systems in wireless communication engineering, particularly focusing on their near-field beam focusing and spatial resolution properties. The term encompasses both the robotics-oriented "X-VLA" architecture for generalist multi-robot learning (Zheng et al., 11 Oct 2025) and the communication theory "XL-array" paradigm for modular, high-resolution antenna design (Li et al., 2023). Both research thrusts involve unifying architectures across heterogeneous physical or data domains, addressing scalability, adaptability, and domain-specific challenges.

1. Architectures for Scalable Vision-Language-Action Models

The X-VLA architecture for scalable VLA modeling is built around a fully shared Transformer backbone for multi-modal fusion, supporting diverse robot embodiments and tasks without requiring architectural branching. Inputs are divided into high-dimensional streams (vision and language, e.g., fixed and wrist-view cameras plus instruction through a frozen Florence-Large encoder) and low-dimensional streams (proprioception and action). Embodiment-specific soft prompts—learnable embeddings $p_i$ for each data source $i$—are prepended at every Transformer layer to embed embodiment-awareness into internal representations. The action generation process employs a continuous-time flow-matching policy, denoising initial Gaussian action chunks toward target trajectories via a velocity field $v_\theta(A^t, o, t)$ integrated using methods such as Euler–Maruyama.

For heterogeneous datasets $\mathcal{D}_1, \ldots, \mathcal{D}_H$ collected on hardware $h_i$, prompts $p_i$ (dimension $\mathbb{R}^k$, $k\approx128$) are optimized end-to-end to absorb platform-specific variation. This allows the backbone to remain fully shared and simplifies scaling to additional platforms and domains (Zheng et al., 11 Oct 2025).

2. Modular XL-Array Theory for Communication Systems

The modular extremely large-scale array (modular XL-ULA) paradigm models antenna arrays as sub-arrayed structures, introducing a detailed non-uniform spherical-wave (NUSW) model for precise amplitude and phase characterization across physically extensive arrays (Li et al., 2023). Each array comprises $N$ modules with $M$ elements per module, module spacing $i$0, and element spacing $i$1, yielding a total aperture $i$2. The user (or scatterer) is parameterized by location $i$3, and the exact element-to-user distance is

$i$4

Two sub-array uniform spherical wave (USW) approximations are derived for tractable analysis:

• USW with different angles: Local plane-wave approximation within each sub-array, allowing module-specific incidence angles.
• USW with common angle: Effective when the user is in the extended far-field of each module, enabling a Kronecker-structured array response.

These models enable explicit, closed-form characterization of beam focusing, spatial/angular resolution, and grating lobe structure.

3. Training Objectives and Data Heterogeneity in X-VLA

X-VLA models are trained using a combination of standard behavior cloning (BC) and a flow-matching BC (FM-BC) objective:

• BC loss: $i$5
• FM-BC loss: For $i$6, $i$7,

$i$8

Empirically, no additional auxiliary losses are employed; domain-specific adapters or hard prompt tuning (HPT) layers were outperformed by the soft prompt approach. Pretraining utilizes 290K episodes from seven robotic hardware variants with distinct perceptual and kinematic properties. Datasets include AGIBOT-beta, Droid, RoboMind, and AgileX platforms with substantial variation in arm kinematics, vision configuration, and control interfaces (Zheng et al., 11 Oct 2025). This design stabilizes multi-domain training and achieves significantly reduced validation error (0.041 for X-VLA-0.9B vs. 0.056–0.14 for alternative architectures).

4. Beam Focusing, Resolution, and Grating Lobes in Modular XL-ULA

For modular XL-ULAs, spatial/angular resolution and sidelobe/grating lobe characteristics are governed by array structure and signal processing regime:

• Angular resolution: Modular arrays achieve narrower main-lobe in $i$9:
  • Modular: $v_\theta(A^t, o, t)$0
  • Collocated: $v_\theta(A^t, o, t)$1
  • where $v_\theta(A^t, o, t)$2.
• Range resolution: For fixed $v_\theta(A^t, o, t)$3, the main-lobe in $v_\theta(A^t, o, t)$4 roughly scales as $v_\theta(A^t, o, t)$5 for modular, superior to collocated arrays' $v_\theta(A^t, o, t)$6.
• Grating lobes: Far-field uniform plane wave (UPW) analysis predicts grating lobes at $v_\theta(A^t, o, t)$7 with full main-lobe amplitude when $v_\theta(A^t, o, t)$8. Under near-field USW, the non-linear phase curvature $v_\theta(A^t, o, t)$9 across modules breaks periodicity, suppressing grating lobes and producing Fresnel-type sidelobe decay.
• Simulation evidence: Validates theoretical predictions on main-lobe width, grating lobe location and suppression, and superior range-focusing in modular arrays (Li et al., 2023).

5. Performance, Generalization, and Empirical Results

The X-VLA-0.9B model (0.9B parameters) achieves new state-of-the-art performance on major simulation and real-world robotics benchmarks:

Benchmark	X-VLA-0.9B Score	Max Prior SOTA
Simpler–VM	80.4%	78.0%
Simpler–VA	75.7%	72.7%
Simpler–WidowX	95.8%	71.9%
Libero (avg)	98.1%	97.1%
Calvin	4.43 tasks	4.53
RoboTwin-2.0 Easy	70.0%	34.5%
RoboTwin-2.0 Hard	39.0%	16.4%
VLABench (avg)	51.1%	39.7%
NAVSIM PDMS	87.3	81.7

In the real world, WidowX pick-and-place, AgileX cloth folding, and AIRBOT cloth pick tasks all show superior or matching performance to closed-source and larger-parameter baselines, with 100% success on cloth folding (33 folds/hr, 1200 demos) and AIRBOT achieving 93% success on Libero with only 200 demos using LoRA fine-tuning ($\mathcal{D}_1, \ldots, \mathcal{D}_H$0 of model parameters), comparable to full fine-tuning on much larger models.

Generalization studies demonstrate the efficiency of two-phase prompt-based adaptation, revealing that adapted prompts optimize convergence and final performance compared to random or transferred (frozen) prompts. Multi-domain finetuning is empirically validated to be as effective or superior to single-domain specialization, evidencing cross-embodiment transfer.

6. Limitations and Future Outlook

Current XL-VLA instantiations are constrained by both model and data scale compared to leading LLM/VLM systems. More diverse and higher-quality robotics corpora, larger backbones, and stronger vision-LLMs are expected to further advance generalization. A known limitation is the sparsity of low-dimensional action labels; incorporating richer supervision (3D spatial cues, physics-informed priors, or subgoal annotations) is likely to improve internal representations. Truly plug-and-play transfer—requiring zero or few demonstrations when adapting to novel embodiments—remains an open challenge; unified kinematics and physics-based abstractions could mitigate the necessity of platform-specific soft prompts. In the communications context, XL-array design must continue to balance the trade-off between spatial resolution and grating lobe suppression, capitalizing on near-field phenomena for deployment in large-aperture, user-dense environments (Zheng et al., 11 Oct 2025, Li et al., 2023).

7. Significance and Theoretical Foundations

XL-VLA research unifies large-scale, data-driven approaches for both embodied robotics and communication systems by leveraging architectures and mathematical analyses that support scalable, cross-domain transfer with minimal intervention. For robotics, soft-prompted Transformers with flow-matching policy enable scalable, generalist robot learning across a diversity of platforms, simplifying adaptation via small parameter-sets while stabilizing multi-domain representation learning. For wireless communications, modular XL-arrays with spherical-wave models offer tractable tools to design systems with high angular and range resolution, exploiting non-linear near-field phase characteristics to suppress undesirable side-lobes. These methodologies collectively set new performance standards in their respective domains, while revealing foundational principles for scalable, generalist system design (Zheng et al., 11 Oct 2025, Li et al., 2023).