Skip to content

论文HF Daily Papers · arXiv 2611.226665 天前

VSpark加速VLA推理:无损解码提速1.28至2.10倍 ​

本文由 AI 根据公开内容整理,可能有误,请以原文为准:VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models(HF Daily Papers · arXiv 2611.22666)

核心内容 ​

针对自回归视觉语言动作(VLA)模型逐Token生成导致机器人控制延迟高的问题,研究提出VSpark推测解码框架。该方法旨在保持解码无损的前提下提升推理速度,在四种采用不同动作Tokenizer的VLA模型及四个操作任务基准测试中,实测取得1.28倍至2.10倍的加速效果,有效缓解了具身智能实时控制中的计算瓶颈。

VSpark包含三项核心设计:一是使用校准动作噪声收集Rollouts,扩展Draft模型对机器人状态的覆盖;二是特征对齐训练,使Draft模型匹配VLA隐藏层特征而非仅拟合输出Token,解决动作语义近似导致的验证失败;三是续用机制(Continuation Reuse),保留被拒绝Draft路径上已计算的预测作为下一步候选,避免算力浪费。该方案从算法层面优化了VLA端侧部署效率。

关键看点 ​

  • 无损加速指标:在4种VLA与4个操作基准上实现1.28x-2.10x推理提速,且保持解码结果与原模型完全一致
  • 特征对齐机制:训练Draft模型对齐VLA隐藏特征而非仅输出Token,解决机器人动作Token语义冗余导致的验证拒绝问题
  • 噪声Rollout策略:利用校准动作噪声采集数据,迫使轻量级Draft模型学习更广泛的机器人状态分布以提升提案质量
  • 计算续用优化:复用被拒绝Draft路径上已计算的预测值作为下一验证步候选,显著减少无效前向传播计算开销

工程与产业观察 ​

VSpark证明了在不牺牲精度的前提下,通过算法级推测解码优化可显著降低VLA推理延迟。其特征对齐与续用机制为具身大模型端侧部署提供了关键工程参考,表明解决动作空间语义冗余是提升机器人实时响应能力的核心技术路径,对高频操作任务落地具有直接价值。

英文摘要 (Abstract) ​

Our aim is to accelerate autoregressive vision-language-action models (VLAs), which generate robot actions one token at a time. Speculative decoding makes this possible: a lightweight draft model proposes several tokens at once, and the VLA verifies them in a single pass, accepting only those that match its own predictions, so decoding remains lossless. The speedup, however, is limited in robot control, where many different tokens encode nearly the same motion and a draft close to the correct action is still rejected. To address this, our approach, VSpark, is built upon three design choices: (i) rollouts collected with calibrated action noise, which expose the draft to a wider range of robot states; (ii) feature alignment, which trains the draft to match the VLA's hidden features rather than its output tokens alone; and (iii) continuation reuse, which keeps the predictions the VLA has already computed along rejected drafts as candidates for the next verification step, instead of discarding them. On four VLAs with different action tokenizers and four manipulation benchmarks, VSpark achieves 1.28x to 2.10x speedup while keeping the VLA's decoding lossless.

Hugging Face Daily Papers 热度:7 upvotes 作者:suho yoo, Youngjoon Jang, Hyebin Cho, Taeyeop Lee, Jaewook Jeong, Joon Son Chung

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 2611.22666:VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models ↗访问 2026-11-20