论文记录 · 证据边界公开

Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos

arXiv technical report; project team reports ICML 2026 acceptance, proceedings unverified · 2025-07-21 · 人类视频到机器人动作

论文元数据

作者: Hao Luo、Yicheng Feng、Wanpeng Zhang、Sipeng Zheng、Ye Wang、Haoqi Yuan、Jiazheng Liu、Chaoyi Xu、Qin Jin、Zongqing Lu
机构: 机构尚未补齐
发布日期精度: 日
研究方向: 人类视频到机器人动作
机器人: 未关联具体机器人
DOI: 10.48550/arXiv.2507.15597
arXiv: 2507.15597
引用元数据: 0
记录状态: included
质量检查: 无自动质量红旗

工程资源状态

代码: 已核验可访问（使用限制见许可证）
权重: 已核验开放
数据: 已核验开放
许可证: 许可证条款已核验（不等于可商用）

已终审工程资源

官方代码 · 许可证 MIT · 最近核验 2026-07-19
依赖：Linux x86_64、Python 3.10.18、PyTorch 2.7.1、transformers 4.43.4、accelerate 1.9.0、bitsandbytes 0.46.0、DeepSpeed 0.15.4、FlashAttention、MANO assets under a separate license；硬件/传感器：GPU required by the documented command; minimum VRAM not disclosed、paper setup: Franka Research 3, Inspire hand and RealSense L515；实机证据：The paper and official videos provide quantitative Franka plus Inspire-hand evidence; independent reproduction was not established.；开放范围/限制：Inference/evaluation and tokenizer code are public, but training scripts, the complete real-robot development stack and simulation benchmark remain unreleased.
官方权重 · 许可证 MIT · 最近核验 2026-07-19
依赖：Being-H0 inference code、GPU; model-specific minimum VRAM not disclosed、MANO assets for the motion pipeline under a separate license；硬件/传感器：paper setup: Franka Research 3, Inspire hand and RealSense L515；实机证据：The paper reports real-robot task success comparisons and the official project hosts real-robot videos; no independent reproduction was verified.；开放范围/限制：The collection contains five verified public model repositories. Open weights do not imply that the unreleased training code or full pretraining data are available.
官方数据集 · 许可证 MIT · 最近核验 2026-07-19
依赖：Being-H0 data loader、upstream test datasets are not redistributed、MANO assets require separate registration and license；硬件/传感器：paper data collection: Franka Research 3, Inspire hand, Gello exoskeleton and RealSense D435i；实机证据：The paper reports 50 to 100 teleoperation trajectories per downstream task; the released object is only the post-training subset.；开放范围/限制：Verified public dataset contains 10,820 files and 1,169,256,624 bytes. It is not the complete UniHand-2.5M pretraining corpus.

“官方”表示归属与论文关系已核验；仍应按所列依赖、硬件和许可证复现。

原始来源

待核验工程线索

模型候选 exact_identifier · discovery_candidate

候选只通过标识符或规范化标题匹配；尚未证明官方归属、可下载、许可证、依赖或实机复现。

最近核验：2026-07-19。链接存在不等于官方仓库、可运行、已开放权重/数据或许可证允许使用。