VLA-Bridge: A Parameter-Efficient Adaptation Framework for Lightweight Vision-Language-Action Modeling
Journal of Innovative Research and Technology · 20.04.2026 · 3 (1) · 70-83
DOI: 10.66857/jirt.v3i1.28 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Innovative Research and Technology, online erschienen am 20.04.2026. Verfasst von Michael Reynolds und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The integration of visual perception, natural language understanding, and continuous motor control has led to the emergence of Vision-Language-Action models. These models demonstrate unprecedented generalization capabilities in embodied artificial intelligence and robotic manipulation tasks. However, extending these architectures to physical edge devices is severely constrained by their immense parameter counts, exorbitant memory footprints, and high latency during real-time inference. Parameter-efficient fine-tuning methods have shown promise in language and vision domains, but they often fail to capture the complex, cross-modal alignments required for executing precise physical actions. In this paper, we propose VLA-Bridge, a novel parameter-efficient adaptation framework specifically designed for lightweight Vision-Language-Action modeling. VLA-Bridge introduces a dynamic cross-modal adapter module that projects high-dimensional visual tokens and linguistic instructions into a compact, shared action-latent space using low-rank approximations. This approach drastically reduces the number of trainable parameters while maintaining the representational capacity necessary for complex robotic control. We conduct extensive experiments on established embodied benchmarks, demonstrating that VLA-Bridge achieves performance on par with full fine-tuning methods while reducing trainable parameters by two orders of magnitude and decreasing inference latency by a significant margin. The proposed framework provides a critical stepping stone toward deploying advanced reasoning and control policies directly on resource-constrained robotic hardware.
Zitieren
APA
Reynolds, M., & Sato, N. (2026). VLA-Bridge: A Parameter-Efficient Adaptation Framework for Lightweight Vision-Language-Action Modeling. Journal of Innovative Research and Technology, 3(1), 70-83. https://doi.org/10.66857/jirt.v3i1.28
Vancouver
Reynolds M, Sato N. VLA-Bridge: A Parameter-Efficient Adaptation Framework for Lightweight Vision-Language-Action Modeling. Journal of Innovative Research and Technology. 2026;3(1):70-83. doi:10.66857/jirt.v3i1.28
BibTeX
@article{Reynolds2026,
author = {Reynolds, Michael and Sato, Noah},
title = {VLA-Bridge: A Parameter-Efficient Adaptation Framework for Lightweight Vision-Language-Action Modeling},
journal = {Journal of Innovative Research and Technology},
year = {2026},
volume = {3},
number = {1},
pages = {70--83},
doi = {10.66857/jirt.v3i1.28}
}
EUVIMED