Connecting real-world tasks
with usable data.
连接真实场景
交付可用数据

First-person stereo human demonstrations for humanoid robotics, VLA and world models. Hardware, sites, collection, data engineering and QC — individually or combined.面向人形机器人、VLA 与世界模型的第一视角双目人类演示数据。设备、场景、采集、数据工程与质检,可单项合作,也可组合实施。

01 / Where we sit01 / 我们的位置

Buyers aren't missing hardware. They're missing data they can use.客户缺的不是设备,是能直接用的数据。

Data that reaches a training pipeline needs five things at once. Every missing one lands back on the buyer as project management.能进训练管线的数据要同时凑齐五件事。少一件,成本就回到客户身上。

Supplier供给方 Hardware设备 Sites场景 Collection采集执行 Data eng.数据工程 QC evidence质量证据
Vision / sensor vendors视觉 / 传感器厂商
Collection-labour outsourcers采集人力外包
10K HoursData
Provided提供 Partial部分 Left to the buyer客户自理
Define hardware, scene and data-service scope around your task.围绕具体任务,确定设备、场景与数据服务边界。
02 / Verifiable output02 / 可核验的产出

Don't read the deck. Decode the sample.别看 PPT,把样片解码看。

Public sample frame — desktop PC internal wiring, first-person stereo
EP001PC wiring · 121.57 s台式机接线 · 121.57 s
Public sample frame — retail restocking, first-person stereo
EP002Restocking · 41.90 s门店理货 · 41.90 s
tatezhou/ego-stereo-cn-v1 CC BY-NC 4.0 passes pi-data-sharing通过 pi-data-sharing LeRobotDataset() Actual capture frames from the published sample上图为已发布样片的真实采集画面
0
Frames · 2 episodes · 163.47 s帧 · 2 集 · 163.47 秒
1920×1200
Per eye, two colour streams每目分辨率,两路彩色
p99 = med
33.29 ms interval · 0 out of spec帧间隔 33.29 ms · 0 超标
0Hz
IMU · 6-DoF head pose per frameIMU · 每帧 6-DoF 头部位姿
03 / Time base03 / 时间基准

Timestamps are recorded, not computed.时间戳是记录出来的,不是算出来的。

IMU and image share one exposure time base — 10 inertial samples anchored inside each frame's exposure window. No back-filling, no two-clock alignment after the fact.IMU 与图像共用同一曝光时基,每帧曝光窗口内锚定 10 条惯性样本。不做恒定帧率回填,不做双时钟事后对齐。

Frame interval帧间隔
0ms
Median, p99 and max all identical中位 / P99 / max 三者相同
Out of spec超标帧
0
Across the full public sample全部公开样片范围内
Timestamp coverage时间戳覆盖
0%
Per frame, no interpolation逐帧,无插值
IMU gaps >10 msIMU >10 ms 空洞
0
18,050 continuous samples18,050 样本连续
04 / Capture hardware04 / 采集硬件

Built for first-person manipulation data.为第一视角操作数据而设计。

Head unit plus belt controller, lenses tilted forward so a natural gaze covers both hands. We define the spec and the data path; manufacturing partners build and deliver.头戴单元 + 便携主控,镜头前倾,平视即覆盖双手。我方定义整机规格与数据链路,联合制造伙伴完成生产交付。

Schematic · not to scale结构示意 · 非比例 head unit 60.7 mm baseline stereo · 2000×1200 · 30 fps IMU · 6-axis · 300 Hz hands + work surface within natural gaze controller · local storage exposure-locked no cloud binding
Sensor传感器
Passive stereo被动双目 · one RGB stream per eye· 每目一路彩色
Resolution分辨率
2000 × 1200 / eye
Frame rate帧率
30 fps · hardware-synced
IMU
300 Hz (measured 300.50)
Field of view视场
VFOV ≈ 90° · HFOV ≥ 115°
Exposure曝光
7.508 ms lockable, constant可锁定,实测恒定
Baseline基线
60.7 mm (60.66 / 60.76)
Calibration立体标定
0.166 px Sampson median
Reprojection重投影
0.800 px stereo · 15,811 inliers
Shipped with data随数据交付
K · KB4 · R1/R2/P1/P2/Q · IMU↔cam · version
Customisable可定制 Form & mounting佩戴与结构 Sensor set传感器组合 Controller & I/O主控与接口 Data format数据格式 Units / modules / OEM整机 / 模组 / OEM
05 / Sites & tasks05 / 场景与任务

Organised by task, across the work people actually do.按任务组织,覆盖人真正在做的作业。

The core team previously built and ran 93 offline locations from zero — site acquisition, staffing, SOPs, cross-city scheduling, exception handling.核心团队曾从 0 组织并运营 93 个线下点位:点位拓展、人员组织、标准化 SOP、跨地调度与异常处理。

Task taxonomy — illustrative references, not capture footage任务类型示意图 — 非我方采集画面 9 families · 36 task types9 类 · 36 种任务

Above: the task families we scope projects against — labs, clinics and pharmacies, training rooms, pet care, light office and logistics, retail, kitchens, repair and electronics, textile and craft. Images are generated references used to describe scope; the only real capture frames on this site are the two published sample frames above.以上为我们对项目范围的任务分类:实验室、诊所与药房、培训、宠物洗护、轻办公与物流、零售、后厨、维修与电子、纺织与手作。图片为用于说明范围的生成参考图;本站唯一的真实采集画面是上文两帧公开样片。

S/01

Site authorisation场地授权

Venue, staff and third-party consent before anything is recorded场地、人员与第三方授权先于任何采集

S/02

Operator training人员训练

Task cards, framing rules, what counts as a rejected take任务卡、构图规则、什么算无效片段

S/03

Devices & logs设备与日志

Per-unit checks, per-session logs, serials bound to batches逐台检查、逐场日志、序列号绑定批次

S/04

Execute & spot-check执行与抽检

Live supervision against the buyer's rejection criteria按买方拒收规则现场盯守与抽检

S/05

Daily close & rework日结与返工

Same-day QC; rework queued before the next shift当日质检,返工在下一班前排队

06 / Delivery & QC06 / 交付与质检

From your schema to a package — not a pile of files.交出去的是数据包,不是一堆文件。

Your fields and acceptance criteria become things a crew can do on site. We close the loop on a pilot sample before batch execution starts.把字段和验收标准翻译成现场做得到的动作。先在小样阶段闭合口径,再进入批量执行。

01

You define客户定义

Schema fields, task types, data format, acceptance criteria and consent requirements.Schema 字段、任务类型、数据格式、验收标准与授权要求。

schematasksthresholdsconsent
02

We translate我方转换

Field-to-source mapping, capture rules, task cards, automated QC thresholds, rework and versioning rules.字段与数据源映射、采集规则、任务卡、自动质检阈值、返工与版本规则。

mappingtask cardsQC rulesSOP
03

You receive客户收到

A pilot sample and QC report first, then standardised packages with manifest and checksums.先收到小样与质检报告,再按批交付标准化数据包、清单与校验文件。

pilotpackagemanifestSHA-256
Formats交付格式 MCAPLeRobot v2.1 / v3Custom schema客户自定义 Schema Every batch每批附 Metadata元数据Auto QC report自动质检报告Review log人工复核记录SHA-256
Acceptance gates · illustrative readout验收门槛 · 示意读数 0/5 gates armed项门槛就绪
  1. Q/01

    Frame rate & timing帧率与时序

    Back-computed from device timestamps; p99 and max reported. No unflagged gap >100 ms within any hour.设备时间戳反算,报 P99 与 max;1 小时内无「未标记丢帧且间隔 >100 ms」的空洞。

    gap ≤ 100 ms
  2. Q/02

    Resolution, colour & IMU分辨率、色彩与 IMU

    Two colour RGB streams at 2000×1200 each, confirmed by sampling. IMU mean and minimum Hz reported; minimum must exceed 200 Hz.抽帧确认两路彩色 RGB、单路 2000×1200;报 IMU 平均与最低 Hz,最低须 >200 Hz。

    IMU > 200 Hz
  3. Q/03

    Sync & completeness同步与完整性

    L/R and IMU-to-nearest-frame deltas as measured distributions. Video, timestamps, IMU and calibration — missing one fails the batch.左右目、IMU 与最近帧时间差给实测分布;视频、时间戳、IMU、标定缺一不合格。

    4/4 streams路齐全
  4. Q/04

    Consent & data sovereignty授权与数据主权

    Venue and staff authorisation, minimum-necessary capture. No cloud binding — raw data stays on local storage end to end.场地与人员授权、最小必要采集;不绑云,原始数据全程本地存储。

    local only全程本地
  5. Q/05

    Per-unit bench test每台出厂检测

    Measured resolution, interval p99/max, IMU rates, sync error distributions, calibration bound to serial, firmware version.分辨率、帧间隔 P99/max、IMU 采样率、同步误差实测分布、标定绑定序列号、固件版本。

    per serial逐台绑定
manifestauto QC report自动质检报告review log人工复核记录SHA-256
Optional add-on可选加购

Stop right at the entrance of the training pipeline.数据可以直接停在训练管线的入口。

Hand 3D keypoints in metric MANO-15, left and right tracked separately, with human-labelled semantic segments. Wrist cameras optional, dual 1920×1080 @ 30 fps.手部 3D 关键点(MANO-15 米制坐标),左右手分轨,语义分段人工标注。腕部相机选配,双路 1920×1080 @ 30 fps。

97.6% L valid97.4% R valid
07 / About07 / 关于我们

A team that has run sites, built pipelines and shipped hardware.一个跑过现场、搭过管线、交付过硬件的团队。

10K HoursData is the brand of 一万小时文化咨询(南京)有限公司, based in Hangzhou and Nanjing.10K HoursData 是一万小时文化咨询(南京)有限公司的品牌,团队 Base 杭州与南京。

93 · offline sites · organised from zero线下点位 · 从 0 组织
0sites点位
A/01

Hardware & data path硬件与数据链路

Full-unit spec, calibration, time base, firmware acceptance with manufacturing partners整机规格、标定、时基、与制造伙伴的固件验收

A/02

Sites & field operations场景与现场运营

Site acquisition, consent, crew training, SOPs, cross-city scheduling点位拓展、授权、人员训练、SOP、跨地调度

A/03

Data engineering数据工程

Schema mapping, LeRobot / MCAP packaging, automated QC and validator runsSchema 映射、LeRobot / MCAP 封装、自动质检与校验器复跑

A/04

Buyer interface客户接口

One technical contact from pilot scoping to batch acceptance从试采范围到批量验收,一个技术对接人

Track record履历
0
Offline locations organised from zero从 0 组织的线下点位
Capacity产能
0+
Accepted hours per month月交付有效小时
Cities城市
0
Hangzhou · Nanjing杭州 · 南京
Backed by投资孵化
Nanjing Zijin Venture Capital南京紫金创投
Team from Zhejiang University · UCLA · Columbia团队来自浙江大学 · UCLA · 哥伦比亚大学
08 / Next step08 / 合作方式

Start with a small batch that's real and acceptable.从一批真实、可验收的小规模任务开始。

We don't promise device counts or hours a pilot hasn't validated. The first batch produces the capacity and cost model.不承诺未经小批量验证的设备数与小时数。首批产出产能与成本模型。

  1. 01

    Minimum start明确最低启动要求

    Schema, task types, acceptance thresholdsSchema、任务类型、验收阈值

  2. 02

    Pilot scope确认小批量范围

    Sites, operators, hours, deadline场景、人员、时长、交期

  3. 03

    Pilot & acceptance执行试采与验收

    Collect, QC, deliver, get your verdict采集、质检、交付、拿到你的验收结论

  4. 04

    Capacity & cost model形成产能与成本模型

    Built from measured pilot numbers:用首批实测数字算:

    concurrent devices并发设备数accepted hours验收有效时长first-pass QC rate一次质检通过率rework causes返工原因cost / accepted hour每有效小时成本
  5. 05

    Scale按质量与节奏扩容

    Weekly and monthly capacity you can hold us to可兑现的周度 / 月度产能

Send us your schema. We'll return a pilot plan and a sample cut to it.把 Schema 发给我们,我们回一份试采方案和一段按它切好的样片。

LeRobot v3, MCAP or custom. 10,000+ accepted hours per month. Hangzhou · Nanjing.LeRobot v3 / MCAP / 自定义格式。月交付 1 万+ 有效小时。杭州 · 南京。

tate@10khoursdata.com