ScaleVLA ties a 552B VLM backbone to a robot action model
Original title: ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
With 10 demonstrations it beat a 2B backbone trained on 50, and LIBERO-PRO success was 45.1% versus at most 31.3%.
ScaleVLA connects a 552 billion-parameter DeepSeek-V4.1-Flash backbone to a trainable action diffusion transformer and compares vision-language models from 2B to 552B under a shared training setup. With 10 demonstrations it reached 87.8%, versus 85.6% for a 2B backbone trained with 50. LIBERO-PRO success was 45.1%, compared with at most 31.3% for smaller backbones. Real-world charger insertion rose from 4/10 to 9/10, with gains in contact-sensitive alignment, at 216 ms median latency per action chunk.
Unconfirmed: Performance figures are the authors' comparisons under their shared training setup.
- The paper cites GPT-6 Astra as an example of frontier VLMs used for robot control.
- Gains on charger insertion were concentrated in contact-sensitive alignment.
- Median latency was 216 ms per action chunk despite the 552B backbone.
- Vendors
- DeepSeek
- Products
- LIBERO-PRO
From the original article
1 min read at huggingface.coFrontier Vision-Language Models (VLMs) such as GPT-6 Astra have demonstrated remarkable capabilities in robot control, suggesting that broad pretraining provides knowledge valuable for physical action. Yet most Vision-Language-Action (VLA) policies still rely on much smaller VLM backbones, leaving this potential largely untapped.
By Ruixi Yu, Dairu Liu, Xuchuan Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.