ScaleVLA scales robot vision-language-action models to 552B parameters
The ScaleVLA paper connects a 552B-parameter DeepSeek-V4.1-Flash backbone to a trainable action DiT and compares vision-language backbones ranging from 2B to 552B parameters. The authors report that the largest model reached 45.1% success on LIBERO-PRO, improved charger insertion from 4/10 to 9/10 successes, and maintained a 216 ms median latency per action chunk.
The author-reported results suggest that much larger pretrained vision-language representations could improve the sample efficiency, generalization, and precise manipulation of robot policies.
Computer Science > Robotics
arXiv:2611.21607 (cs)
[Submitted on 16 Nov 2026]
Title:ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
Authors:Ruixi Yu, Dairu Liu, Xuchuan Chen, Zekun Qi, Zekai Li, Zhuo Cao, Luyang Niu, Yingsheng Zhang, Sikai Liang, Ke Fan, He Wang, Li Yi
View a PDF of the paper titled ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters, by Ruixi Yu and 11 other authors