ScaleVLA Scales Vision‑Language‑Action Model to 552 B Parameters
training
| Source: HF Papers | Original article
Researchers introduce ScaleVLA, a Vision‑Language‑Action model scaled to 552 billion parameters, aiming to bridge the gap between large pre‑trained VLMs and robot control.
ScaleVLA, a new Vision‑Language‑Action (VLA) model, has been announced with a staggering 552 billion parameters, marking the first time a VLA architecture has been scaled to the size of today’s largest language models. The effort builds on observations that frontier Vision‑Language Models such as GPT‑6 Astra can already drive robot control, hinting that extensive pre‑training endows visual‑language systems with knowledge useful for physical tasks. Until now, most VLA policies have been built on comparatively tiny VLM backbones, limiting their ability to generalise across diverse manipulation scenarios. ScaleVLA closes that gap by marrying a massive visual‑language core with an action‑generation head, enabling a single model to interpret visual input, understand language instructions and output motor commands.
The significance lies in the potential to streamline robot‑learning pipelines. By leveraging a unified, heavily pre‑trained backbone, developers may reduce the need for task‑specific fine‑tuning and large amounts of robot‑specific data, accelerating deployment of autonomous systems in manufacturing, logistics and service sectors. Moreover, the release signals that the research community is now treating VLA as a first‑class scaling target, rather than an afterthought to language‑only models.
Going forward, the community will watch for benchmark results that compare ScaleVLA against existing VLA baselines on standard robot‑control suites. Attention will also turn to the compute and data requirements needed to train and run such a model, and whether more efficient variants or open‑source equivalents will emerge. The rollout could set a new performance ceiling for embodied AI and shape the next wave of robot‑centric applications.
Sources
Back to AIPULSEN