← Paper Digest

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters

Ruixi Yu, Dairu Liu, Xuchuan Chen, Zekun Qi, Zekai Li, Zhuo Cao, Luyang Niu, Yingsheng Zhang, Sikai Liang, Ke Fan, He Wang, Li Yi

November 2026

ScaleVLA on the real robot aligning the charger with the power strip, seen from the top and both wrist cameras.

TL;DR Attach a frozen 552B mixture-of-experts VLM to a trainable action DiT and compare it with 2B-32B backbones under identical training. The 552B policy learns LIBERO-Goal from 10 demonstrations better than a 2B backbone does from 50 (87.8% vs 85.6%) and inserts a charger on a real bimanual robot in 9/10 trials vs 4/10, at 216 ms per action chunk.

What they do

Most VLAs build on VLMs with a few billion parameters. ScaleVLA tests whether a frontier-scale VLM helps: it attaches the frozen 552B mixture-of-experts DeepSeek-V4.1-Flash to an action head trained from scratch, and compares it with 2B-32B Qwen3-VL and Cosmos-Reason2 backbones under the same action head, initialization, data and training budget.

How it works

  • Backbone readout. Only the VLM’s 20-layer causal encoder runs, once per action chunk, with no text generation. Its image and text tokens go to the action head.
  • Action head. A GR00T N1.7-style feature adapter, state and action encoders, and a 32-layer DiT alternating self- and cross-attention, trained with flow matching; four flow steps produce each chunk.
  • Training and serving. Gradients reach only the action head, so backbone features are precomputed once. Inference runs on four H200 GPUs, and the mixture of experts activates only part of the model per token.

Results

  • LIBERO average 96.1% vs 90.7% for Cosmos-Reason2-2B. With 5 demonstrations per task, LIBERO-Goal success rises from 39.4% (2B) to 77.4%.
  • Generalization: 45.1% under LIBERO-PRO perturbations vs at most 31.3%; 54.9% in out-of-distribution RoboCasa kitchens vs 42.8-45.6% (100 demonstrations); 42.1% on DexJoCo bimanual tasks vs 31.9-38.1%.
  • Real charger insertion: 9/10 in-domain and 8/15 OOD, vs 4/10 and 5/15 for the 2B backbone and 8/10 and 7/15 for a robot-pretrained π0.5 backbone. All in-domain failures happen at insertion, which ScaleVLA finishes fastest (median 58.1 s vs 81.2 s and 85.9 s).
  • Latency: 216 ms per chunk vs 115 ms for the 2B backbone, with 276x the parameters; 384 ms round trip on the robot.

Limitations

ScaleVLA needs four H200 GPUs where smaller backbones use one, and it receives more input tokens per observation, an effect on latency the authors did not isolate. Each policy uses one training seed, and the real-robot evaluation is a single task. With longer training the 2B backbone narrows the low-data gap, so the fixed-budget sample-efficiency results are not a comparison at convergence.

Videos

A second real-robot Plug In Charger trial, from handover to insertion.
A third Plug In Charger trial on the real robot.

Figures

From Figure 1: ScaleVLA overview: backbone sizes, the VLM feeding an action DiT, task success by backbone size, and R...
From Figure 1: ScaleVLA overview: backbone sizes, the VLM feeding an action DiT, task success by backbone size, and RoboCasa latency, LIBERO-Goal sample efficiency and LIBERO / LIBERO-PRO success.
From Figure 2: The VLM encodes the images and instruction, and the action head reads the output of its causal encoder...
From Figure 2: The VLM encodes the images and instruction, and the action head reads the output of its causal encoder. The action DiT alternates cross- and self-attention and generates an action chunk in four flow steps.
From Figure 5: Plug In Charger on a real robot. All three policies grasp and hand over the charger; ScaleVLA inserts ...
From Figure 5: Plug In Charger on a real robot. All three policies grasp and hand over the charger; ScaleVLA inserts the plug while the other two reach the socket but fail to insert. Right: in-domain and OOD success rates.
Abstract

Frontier 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. We ask whether frontier-scale representations can provide a stronger foundation for robot learning. We introduce ScaleVLA, which connects a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, and systematically compare VLMs from 2B to 552B under a shared training setup. We evaluate three fundamental capabilities of robot foundation models. In sample efficiency, just 10 demonstrations outperform a 2B backbone trained with 50 (87.8% vs. 85.6%). In generalization, LIBERO-PRO success reaches 45.1%, compared with at most 31.3% for smaller backbones. In precise manipulation, real-world charger insertion improves from 4/10 to 9/10 successes, with gains concentrated in contact-sensitive alignment. Despite its scale, ScaleVLA achieves 216 ms median latency per action chunk. Our findings suggest that frontier VLMs can serve not only as reasoning agents, but also as powerful representational foundations for efficient and generalizable robot policies.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.