---
格式版本: 2
标题: "Hydra-0: Action Flow for Generalist World Modeling and Control"
原文链接: "https://arxiv.org/abs/2608.18077"
发布日期: "2026-08-18"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Tue, 18 Aug 2026 17:59:30 UTC (12,559 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T14:34:07+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-20T14:32:53+08:00"
入库时间: "2026-08-20T06:34:07.116Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=NVIDIA&searchtype=all"
匹配关键词:
  []
相关厂家:
  - "NVIDIA"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 9
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该论文为机器人世界模型与控制研究，与超节点/AI Rack/机柜级AI基础设施完全无关，仅来源于arXiv且无相关技术或商业信号。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T14:34:31+08:00"
AI主题相关性: 0
AI来源权威性: 2
AI新颖性: 2
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 5
AI摘要: "Hydra-0 是一种以动作流为条件的通用世界模型，将机器人动作表示为像素运动，从而跨本体、任务和环境实现通用建模与控制。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:21:07.162Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.18077"
---

## Computer Science > Robotics

## Title:Hydra-0: Action Flow for Generalist World Modeling and Control

[View PDF](https://arxiv.org/pdf/2608.18077) [HTML (experimental)](https://arxiv.org/html/2608.18077v1)

> Abstract:We introduce Hydra-0, a generalist world model conditioned on action flow, which represents robot actions as pixel motion. This shared visual interface enables generalist world modeling and control by learning action consequences across embodiments, tasks, environments, and video-generation backbones. Our best configuration achieves 90.4% lower robot-motion error and 60.2% lower object-motion error than our action-conditioned baseline, while supporting zero-shot composition and data-efficient adaptation. On the RoboLab benchmark, Hydra-0 achieves a Pearson correlation of r=0.96 between replayed and reference success rates. Finally, we uncover an emergent inverse mode of this interface: a world action model that predicts compatible robot motion from desired object flow transferred from a human demonstration. A trained action head maps the resulting latent features to executable actions without requiring task-specific expert robot demonstrations. Together, these results demonstrate the potential of action flow as a shared control interface connecting heterogeneous training data, open-loop policy evaluation, and robot control.

| Comments: |  |
| --- | --- |
| Subjects: | Robotics (cs.RO) |
| Cite as: | [arXiv:2608.18077](https://arxiv.org/abs/2608.18077) \[cs.RO\] |
|  | (or [arXiv:2608.18077v1](https://arxiv.org/abs/2608.18077v1) \[cs.RO\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.18077](https://doi.org/10.48550/arXiv.2608.18077) |

## Submission history

From: Hongyu Li \[[view email](https://arxiv.org/show-email/2173fa24/2608.18077)\]  
**\[v1\]** Tue, 18 Aug 2026 17:59:30 UTC (12,559 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.18077) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))