Scaling laws and embodiment-variability interaction in VLA models

Determine the scaling laws that govern the performance and generalization of Vision-Language-Action (VLA) models as model size, data diversity, and data volume increase, and ascertain how embodiment-specific variability in hardware configurations interacts with model capacity during large-scale training and deployment.

Background

The paper presents X-VLA, a soft-prompted transformer framework for cross-embodiment vision-language-action modeling, and demonstrates consistent scaling trends with increased model capacity, data diversity, and data volume. However, the authors note that even their largest tested configuration shows no sign of saturation, suggesting further gains are possible with additional scaling.

In discussing limitations, the authors highlight that computational constraints and the limited availability of high-quality robotics data currently restrict scaling. They explicitly point out that extending X-VLA to larger capacities and broader datasets raises open questions regarding the scaling laws of VLA models and the role of embodiment-specific variability (e.g., differences in hardware configurations) in determining how model capacity translates into performance.

References

Such extensions also raise open questions about the scaling laws of VLA models and how embodiment-specific variability interacts with model capacity.

— X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model  (2510.10274 - Zheng et al., 11 Oct 2025) in Appendix, Section "Limitations and future works"

A related debate in Science Robotics also treats the roles of data, models, and engineering in robotics as an open question.

— The Embodiment Gap in Robot Foundation Models  (2608.18433 - Domae et al., 19 Aug 2026) in Section 1, Introduction

The field has not yet established which data should be collected or which representations should be scaled for that goal.

— The Embodiment Gap in Robot Foundation Models  (2608.18433 - Domae et al., 19 Aug 2026) in Section 3.3, Three scaling directions and why the gap remains

Consequently, the present results do not establish how UVTA behaves with substantially larger and more diverse human datasets, larger models, or new robot embodiments.

— Unified Visual-Tactile-Action Modeling from Human Demonstrations for Dexterous Manipulation  (2609.34182 - Li et al., 28 Sep 2026) in Appendix, Section “Limitations and Future Work,” subsection “Data and model scale”

Joint tokenization of embodiment, state, and action also supports robot co-design and cross-embodiment control, but scalability to humanoid BFM pretraining remains unverified.

— TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments  (2610.17999 - Lan et al., 13 Oct 2026) in Section 2, Related Work, paragraph “Multi-Embodiment Generalist Control”