Noitom Robotics 开放 HiPHI-Hand 数据集,曾用于训练 NVIDIA EgoHand
Noitom Robotics Opens HiPHI-Hand Dataset, Used to Train NVIDIA's EgoHand Model
Noitom 是国内做动作捕捉的老牌厂商,这次把 11 小时带手部真值的第一人称双目视频放到 Hugging Face 上免费开放。做过具身智能或手部姿态估计的朋友都懂这种带 ground truth 的数据有多难采,直接下载就能用来训练,做灵巧手研究的团队可以关注。
Noitom Robotics 在 Hugging Face 上开源了 HiPHI-Hand 数据集,包含 11 小时第一人称视角双目视频,并配有基于标记点的手部动作真值标注。NVIDIA 在训练 EgoHand 模型时将其作为数据来源之一,说明该数据集已经过工业级验证。
数据集采用自我中心(egocentric)双目立体视频形式,配合光学动捕标记获得高精度手部姿态真值,这类数据在灵巧手操作和具身智能研究中长期稀缺。开放后,研究团队可以绕过自建动捕采集流程,直接用手部真值数据训练以自我视角为核心的手部姿态估计与操作模型。
对机器人与具身智能领域来说,第一人称手部交互数据是连接人类演示到机器人灵巧操作的关键环节。HiPHI-Hand 与 NVIDIA EgoHand 的关系也显示头部厂商正在通过开源数据集补齐训练数据短板,中国动捕厂商的采集能力正进入国际主流训练管线。
Noitom Robotics Opens HiPHI-Hand Dataset, Used to Train NVIDIA's EgoHand Model
Noitom Robotics released HiPHI-Hand on Hugging Face: 11 hours of egocentric stereo video with marker-based hand ground truth, which NVIDIA used as one source for EgoHand.
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include HiPHI-Hand, NVIDIA EgoHand, Noitom Robotics, Hugging Face.
Primary source verification: published directly by pandaily with first-party authority.