Skip to content

论文HF Daily Papers · arXiv 2611.216076 天前

ScaleVLA:552B参数VLA模型实现低样本高效操控与216ms推理 ​

本文由 AI 根据公开内容整理,可能有误,请以原文为准:ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters(HF Daily Papers · arXiv 2611.21607)

核心内容 ​

针对现有VLA模型骨干规模受限问题,研究团队推出ScaleVLA,首次将552B参数的DeepSeek-V4.1-Flash前沿VLM作为表征基座,并接入可训练的动作扩散Transformer(DiT)。在统一训练配置下系统对比2B至552B模型性能,证实超大规模预训练知识能有效迁移至物理交互任务,突破了传统小参数VLA在样本效率与泛化性上的瓶颈。

实验数据显示,ScaleVLA仅需10条演示即达87.8%成功率,优于2B模型50条演示的85.6%;LIBERO-PRO基准泛化成功率升至45.1%(小模型最高31.3%);真机充电器插入任务从4/10提升至9/10,增益集中在接触敏感对齐环节。尽管参数量巨大,其中位动作块推理延迟控制在216ms,证明前沿VLM可作为高效通用的机器人策略表征基础。

关键看点 ​

  • 架构创新:采用552B DeepSeek-V4.1-Flash作VLM骨干,耦合可训练DiT动作头,解锁前沿大模型物理表征能力
  • 样本效率:仅10条示教成功率87.8%,超越2B骨干50条示教的85.6%,大幅降低数据采集成本
  • 泛化与实操:LIBERO-PRO成功率45.1%领先小模型14个百分点,真机精密插拔成功率由40%跃升至90%
  • 推理性能:552B参数规模下动作块中位延迟仅216ms,满足实时操控需求,兼顾规模与响应速度

工程与产业观察 ​

ScaleVLA证实552B级VLM不仅擅长推理,更能作为强表征基座显著提升具身策略的样本效率与泛化上限。其216ms延迟表明超大模型经DiT解耦后具备实时部署潜力,为行业探索“大模型+小动作头”范式提供了关键量化依据,推动具身智能从专用小模型向通用大表征演进。

英文摘要 (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.

Hugging Face Daily Papers 热度:39 upvotes 作者:Ruixi Yu, Dairu Liu, Xuchuan Chen, Zekun Qi, Zekai Li, Zhuo Cao, Luyang Niu, Yingsheng Zhang, Sikai Liang, Ke Fan, He Wang, Li Yi

基于本页技术方向关联的整机参数、核心零部件、实测报告与同赛道选型横评。

外部来源

本页事实对应的公开页面,以原文为准。(点击将离开本站)

  1. HF Daily Papers · arXiv 2611.21607:ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters ↗访问 2026-11-18