ScaleVLA Wires 552B DeepSeek-V4.1-Flash Backbone to Robot Policy
TL;DR
- ScaleVLA connects a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT and compares VLM sizes from 2B to 552B under one recipe.
- With just 10 demonstrations the 552B model scores 87.8% versus 85.6% for a 2B backbone trained on 50 demonstrations.
- Real-world charger insertion rises from 4/10 to 9/10 successes at 216 ms median latency per action chunk.
Ten demonstrations were enough. A 552-billion-parameter vision-language-action model called ScaleVLA, trained on 10 examples, beat a 2B backbone trained on 50 (87.8% vs. 85.6%) on the same benchmark, according to a paper posted to Hugging Face this week by researchers at Tsinghua, Peking, Shanghai Jiao Tong and Galbot.
The setup wires a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, then holds that recipe fixed while swapping the VLM from 2B up to 552B. The authors write that 'just 10 demonstrations outperform a 2B backbone trained with 50 (87.8% vs. 85.6%),' framing sample efficiency as the first of three capabilities they probe.
On LIBERO-PRO, a generalization split, success 'reaches 45.1%, compared with at most 31.3% for smaller backbones.' Real-world charger insertion goes from 4/10 to 9/10 successes, with gains the paper says are 'concentrated in contact-sensitive alignment.'
Latency, the usual counter to any scale-up claim, is reported as 216 ms median per action chunk. The paper does not state the hardware producing that figure.
DeepSeek-derived work keeps surfacing on our tracker: 59 DeepSeek stories in the last 90 days, now visibly extending into robotics.
Originally reported by huggingface.co
Read the original article →Original headline: Tsinghua-Led ScaleVLA Scales Vision-Language-Action Models to 552B on DeepSeek-V4.1 Backbone