独立行业技术评审 · 持续核验

ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

ACE-Data-0 是面向世界模型、VLA和模仿学习的人类具身采集计划:官方论文和数据卡描述 150+ 小时、17M+ 视频帧、75,000+ interaction episodes、200+ 任务类别、第一/第三视角视频、全身及手部动作、物体6DoF、音频与全掌触觉。官方 Hugging Face 仓库已建立个人非商业学术研究申请和专门许可,但 README 明示数据文件尚未发布,当前仅有README和LICENSE;因此这是“官方受限发布面已核验、数据制品未开放”的高价值观察记录,不能作为可下载训练集或开放数据规模计数。

ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

ACE-Data-0 是面向世界模型、VLA和模仿学习的人类具身采集计划:官方论文和数据卡描述 150+ 小时、17M+ 视频帧、75,000+ interaction episodes、200+ 任务类别、第一/第三视角视频、全身及手部动作、物体6DoF、音频与全掌触觉。官方 Hugging Face 仓库已建立个人非商业学术研究申请和专门许可,但 README 明示数据文件尚未发布,当前仅有README和LICENSE;因此这是“官方受限发布面已核验、数据制品未开放”的高价值观察记录,不能作为可下载训练集或开放数据规模计数。

开放状态
项目已核验·制品未开放
许可证
ACE-Data-0 Research License Agreement: personal, non-transferable, non-commercial academic research only; data files not yet published
商业使用
not permitted under the published research license; commercial licensing is explicitly handled separately
人工评审结论
official_gated_research_license_metadata_release_data_files_not_published
机器人
human-operated ACE capture rig; no robot-action training embodiment released
任务
atomic household hand-object interaction, long-horizon household activity chains and whole-body human-scene interaction
模态
4 egocentric fisheye RGB views、8 synchronized exocentric RGB views、headset IMU and 6-DoF pose、41-joint body motion、articulated hand motion、SMPL-X、object meshes and 6-DoF trajectories、audio、full-palm tactile pressure、goal and sub-goal language
格式
not released; official README says future data will use sharded archives and will later add layout, checksums and loading examples
采集/生成方式
spatially calibrated, time-synchronized recording in real homes using headset cameras, exocentric cameras, optical motion capture, human motion suit and tactile gloves
已知问题
Data files are explicitly not published; only README/LICENSE metadata tree is public. Access is personal, non-transferable and non-commercial academic only because recordings contain identifiable people in real homes. No manifest, bytes, samples, checksums, loading code, data splits, released calibration, action retargeting, robot embodiment, failure ledger or independent reuse report is available. Do not score planned 150h/75k episodes as downloadable corpus scale.

论文原文 · 官方项目页 · 官方受限数据发布面 · 官方研究许可证