设备 + 整条数据管线一起交付，建成你自己的第一人称数据工厂 Hardware plus the full data pipeline — deploy your own egocentric data factory

一套设备 + 管线，贯通采集、处理、交付到训练Hardware + pipeline in one stack, capture → process → deliver → train

你定义要让机器人学会的技能，我们交付设备与整条数据产线，私有化部署、交钥匙落地。数据是你自建产线跑出来的产出，归你所有、不出域——我们的主业是交付设备与管线能力，而不是按条零售数据。You define the skill; we deliver the hardware and the entire data production line, deployed on-premises, turnkey. Data is what your own pipeline produces — yours, in your domain. We sell hardware and pipeline capability, not data by the clip.

四种方式拥有这条数据产线Four ways to own the pipeline

既可整套交钥匙、私有化部署到你内部，也可只买设备、只授权软件，或联合共建陪跑。设备始终可单独采购，数据始终归你所有。Take the whole stack turnkey on-premises, buy headsets only, license the software, or co-build. Hardware is always sold separately, and the data is always yours.

硬件、管线、场景，三位一体的数据能力Hardware, pipeline and scenarios — one integrated data stack

设备可以单独采购，但更大的价值是把采集终端、处理产线与数据平台整套交付、私有化部署——你拥有的是一条持续产出训练数据的自有产线，数据资产完全归你。Headsets can be bought on their own — but the bigger value is the full stack of capture hardware, processing pipelines and data platform deployed on-premises. You own a production line that keeps generating training data, and you own the data.

Y6 六目采集设备：为数据质量而生Y6 Six-Camera Rig: built for data quality

双目能"看见"，六目才能"测准"。更多灰度相机提供稳定的立体基线与冗余视角，让全天候、长时序作业下的 SLAM 与手部轨迹更稳、更准。Stereo can see; six cameras can measure. Extra grayscale cameras add stable baselines and redundant views, keeping SLAM and hand trajectories robust across long, all-day sessions.

YOUPU Y6 · 6-CAMERA EGOCENTRIC RIG

4 灰度全局快门 + 2 RGB · 后置电池前后配重 · 数据接口全开放、可深度定制

4 grayscale global-shutter + 2 RGB · rear-balanced battery · fully open & customizable data interface

370g

整机重量Weight

<1μs

硬件同步HW sync

6

同步相机Cameras

视觉配置Vision

4 × 灰度全局快门 + 2 × RGB（两种 RGB 版本可选）4 × grayscale global shutter + 2 × RGB (two RGB options)

同步精度Sync

多相机硬件同步 <1μs，帧间隔 ±10msMulti-cam HW sync <1μs, frame interval ±10ms

机内算力On-device

机内直出 SLAM，近距中位误差 5–7mmOn-device SLAM, near-field median 5–7 mm

人体工学Ergonomics

后置电池前后配重，长时佩戴不压头Rear battery balances weight for all-day wear

开放性Openness

数据接口全开放，支持深度定制与私有部署Fully open data interface, deep customization & private deployment

IMU 全温域标定IMU Calibration

• 零偏 / 比例 / 正交误差补偿Bias, scale & orthogonality compensation
• 全温域补偿，8 小时以上标定Full-temperature compensation, 8h+ runs
• IMU–相机联合外参标定Joint IMU–camera extrinsic calibration

相机与立体标定Camera & Stereo

• 内外参重投影误差 0.2 pxIntrinsic/extrinsic reprojection 0.2 px
• 立体校正 0.5 px，鱼眼畸变建模Stereo rectification 0.5 px, fisheye model
• 近距 SLAM 中位误差 5–7mm，RPE 0.67°Near-field SLAM median 5–7 mm, RPE 0.67°

灰度下视 · 稳定追踪Downward Grayscale

灰度全局快门在弱纹理、快速运动、光照变化下仍稳定锁定特征，是长时序不漂移的关键。

Grayscale global shutter locks features under weak texture, fast motion and lighting shifts — key to drift-free long sessions.

双目 / 四目 / 六目怎么选Stereo vs. Quad vs. Six Cameras

我们不与成熟双目/四目厂商对立——管线对主流设备开放兼容；但要全天候、长时序、毫米级的高质量数据，六目是更稳的答案。We don't compete with mature stereo/quad vendors — the pipeline is open to mainstream rigs. But for all-day, long-horizon, millimeter-grade data, six cameras are the robust answer.

真实佩戴，覆盖全场景Real wear across every scenario

戴在额头、双眼露出，无显示无镜片；从实验室、车间到手术室，长时间佩戴依然舒适。Forehead-worn with eyes free, no display or lens; comfortable for long sessions from labs and workshops to operating rooms.

佩戴侧视 · 后置电池Side · rear battery

正面佩戴Front wear

实验室Laboratory

车间Factory

手术室Operating room

KineFlow：从原始流到可训练数据KineFlow: from raw streams to train-ready data

质量门前置、自动结构化、机器质检 + 分层人工抽检 + 每次交付机器审计——把脏活累活做在交付之前，客户拿到的是干净、对齐、带语义的多模态数据。Quality gates up front, automated structuring, machine QC plus tiered human review and a machine audit on every delivery — clean, time-aligned, semantic multimodal data, not raw footage.

多路原始源视频：RGB、灰度、IMU、深度在同一时间轴上硬同步采集。

Raw multi-channel sources: RGB, grayscale, IMU and depth captured on one hardware-synced timeline.

1

原始数据接入与质量门Ingestion & Quality Gate

自动剔除空闲、低质、手部遮挡片段，同步校验、丢帧检测前置。Auto-rejects idle, low-quality and hand-occluded clips; sync and drop-frame checks up front.

2

视频 3D 重建3D Reconstruction

多目立体 + SLAM 恢复相机轨迹、场景几何与真深度流。Multi-view stereo + SLAM recover camera trajectory, scene geometry and true depth.

3

多模态数据提取Multimodal Extraction

手部 / 上半身姿态、3D 轨迹、物体检测分割、IMU、PII 脱敏。Hand/upper-body pose, 3D trajectories, object detection & segmentation, IMU, PII masking.

4

仿真环境融合Simulation Fusion

数字孪生重建与物理参数对齐，支持合成数据与仿真训练。Digital-twin reconstruction and physics alignment for synthetic data and sim training.

5

训练数据交付Training Data Delivery

MCAP 打包：同步 3D 视频 + 真深度 + 元数据 + 时间动作标签。MCAP packages: synced 3D video + true depth + metadata + temporal action labels.

多视角几何与极线对齐Multi-view Geometry

六目硬同步 + 立体校正，输出对齐的多视角与深度。Hardware-synced six views with rectification, aligned multi-view and depth.

手部姿态与 HOIHand Pose & HOI

21 关节手部骨架、人–物交互与接触关系。21-joint hand skeleton, human–object interaction and contacts.

3D 轨迹与身体位姿3D Trajectory & Body Pose

毫米级相机轨迹、上半身关节与速度日志。Millimeter camera trajectory, upper-body joints and velocity logs.

任务语义结构化Task Semantics

活动标签、注意力项、动作分段，输出机器可读 JSON / MCAP。Activity labels, attention items, action segmentation as machine-readable JSON / MCAP.

全身位姿数据产物Full-body Pose Output

在双手之外，管线进一步输出躯干、骨盆与全身骨架：以 SMPL/SMPL-X 参数化蒙皮人体与 MANO 手模表达，关节点、点云、网格与轨迹在 Foxglove 内直接三维预览，全身动作与接触关系一目了然。Beyond the hands, the pipeline outputs the torso, pelvis and full skeleton as SMPL/SMPL-X parametric bodies with MANO hands; joints, point cloud, mesh and trajectory preview in 3D directly inside Foxglove.

SMPL / SMPL-XMANO handsBody jointsPelvis & spinePoint cloudFoxglove 3D

多模态数据层Multimodal Data Layers

每一段数据都可按需叠加以下结构化数据层，构成物理 AI 所需的完整观测。Each clip can be enriched on demand with the structured layers below — the full observation Physical AI needs.

kineflow.studio / dataset / episode_0217.mcap 配套训练平台Companion Platform

MCAP · MULTIMODAL REVIEW

同步通道Channels3D video · depth · IMU

手部姿态Hand pose21 joints

时间标签Action labelsframe-aligned

手部入镜率Hands-in-frameup to 99%

质检QA机器 + 分层人工Machine + tiered human

手部姿态演示Hand Pose Demo 数据清洗演示Cleaning Demo

质量，是可量化的Quality, quantified

100 段真实任务、双目 VST、21 关节手部实测；SLAM 轨迹以机械臂 GT 与 AprilTag-PnP 视觉 GT 为基准。Internal benchmark over 100 real tasks, stereo VST, 21-joint hands; SLAM benchmarked against arm GT and AprilTag-PnP visual GT.

七合一数据平台，从采集到真机评测全程在线Seven products in one console — from capture to real-robot evaluation

不只是硬件和数据：运营驾驶舱、采集任务网络、KineFlow 流水线、多模态标注、数据资产、仿真孪生与真机评测构成完整的数据基础设施，全部流程可观测、可追溯、可重跑。点击下方模块查看各平台界面。Beyond hardware and data: ops cockpit, capture network, the KineFlow pipeline, multimodal annotation, data assets, simulation & twin and real-robot evaluation form a complete, observable, traceable and re-runnable data infrastructure. Click a module to explore each interface.

console.youpu.ai //ops/overview 实时LIVE

‹ ›

采集运营驾驶舱Operations Overview

设备在线、采集吞吐、批次合格率与全球采集网络一屏总览，异常实时告警，让数据产能与质量随时可控。Rigs online, capture throughput, batch pass rate and the global site network on one screen, with live anomaly alerts — capacity and quality always under control.

326

在线采集设备Rigs online

1,840h

当日采集时长Hours captured today

99.2%

批次合格率Batch pass rate

采集任务网络Capture Task Network

跨区域站点任务看板，排期、采集、质检、交付四态统一调度，进度自动回流，支持 24/7 大规模众包采集产线。Cross-region kanban scheduling across scheduled, capturing, QC and delivered states, with automatic progress feedback and 24/7 large-scale capture lines.

42

采集站点 · 9+ 区域Sites · 9+ regions

47

进行中任务Tasks running

2,510

当日采集片段Episodes today

KineFlow 流水线工作台KineFlow Studio

DAG 流水线可视化编排，节点状态、作业运行、自动标注与审计指标实时可见；流水线即代码，失败自动隔离并触发分层人工复核。Visual DAG orchestration with live node status, job runs, auto-labeling and audit metrics. Pipeline-as-code: failures auto-quarantine and trigger tiered human review.

99.1%

自动标注率Auto-label rate

0.8%

隔离复核率Quarantined

100%

交付审计覆盖Audited

多模态标注工作站Annotation Studio

基础模型完成 99% 以上预标注，审核团队只聚焦遮挡、弱光、罕见动作等边界样本；2D 框、双手骨架、3D 点云、HOI 接触与动作分段在同一时间轴对齐。Foundation models pre-label 99%+; reviewers focus on occlusion, low light and rare-action edges. Boxes, two-hand skeletons, 3D point cloud, HOI contacts and segments align on one timeline.

99.1%

预标注帧覆盖Frames pre-labeled

21×2

双手关节骨架Hand joints

5–9mm

关节 3D 不确定性Joint 3D σ

数据资产平台Dataset Hub

真实采集、仿真合成、视频切条统一索引，用自然语言即可检索；每张数据集卡带完整 MCAP Schema、版本与血缘，支持客户专属工作区与私有化部署。Real capture, synthetic data and video cuts share one index, searchable in natural language. Every dataset card carries full MCAP schema, version and lineage, with dedicated workspaces and private deployment.

1.2M

已索引片段Episodes indexed

3-in-1

真实 / 仿真 / 视频Real / sim / video

100%

Schema · 版本 · 血缘Schema · lineage

仿真与数字孪生Simulation & Digital Twin

把真实场景重建为带物理参数的数字孪生，通过光照、材质、相机位姿、姿态重定向与接触力扰动等域随机化，把每段真实数据扩增数十倍，显著缩小 sim2real 差距。Real scenes become physics-enabled digital twins. Lighting, texture, camera pose, retargeting and contact-force randomization expand each clip dozens of times, sharply closing the sim2real gap.

×48

单段真实数据扩增Synthetic per clip

−63%

sim2real 差距Sim2real gap

91.2%

真机一次成功率First-try success

真机评测竞技场Eval Arena

同一公开模型 checkpoint、同一组随机种子，只更换训练数据，在 7 大任务族真机上对比成功率；每格 20 次真机试验，结果可量化、可复现、可审计，让数据价值用真机表现说话。Same public checkpoint and seed set, only the training data changes, benchmarked on real hardware across 7 task families — 20 real trials per cell, quantified, reproducible and auditable. Data value proven by real-robot results.

+34 pts

平均成功率提升Average gain

+51 pts

最大提升 · 医疗精细Largest · surgical

700

本轮真机试验Real trials

datahub.youpu.ai // tools/mcap-reviewREAL · FOXGLOVE EGO DataHub · 具身数据工具 —— 本地 MCAP 回放（Foxglove 官方渲染器）EGO DataHub · MCAP Tool — local MCAP replay (Foxglove renderer) 82 路通道同步回放：多目 RGB、3D 点云、双手 21 关节骨架、深度与 IMU，逐帧质检与时间轴对齐。82 channels replayed in sync: multi-cam RGB, 3D point cloud, 21-joint two-hand skeleton, depth and IMU, with frame-level QC. datahub.youpu.ai // pipelines/registryEGO DATAHUB 算法产线 · 7 条产线与版本管理Pipeline Registry · 7 lines & versioning 手部活动分段、SLAM 建图、机械臂轨迹规划、检测定位、语义分割、点云配准、深度估计，版本 / 任务 / 产物全程可追溯。Hand segmentation, SLAM, trajectory planning, detection, segmentation, point-cloud registration and depth — versions, jobs and artifacts all traceable. harbor.youpu.ai // capture/tasksHARBOR 采集交付 · 三级编目任务库与采集点网络Capture Delivery · 3-level task library & site network 任务按行业 / 场所 / 手持工具三级编目，覆盖家庭、医药仓储、超市、工厂；长三角、珠三角、京津冀采集点实时在线。Tasks catalogued by industry / venue / hand tool across home, pharma storage, supermarket and factory; Yangtze, Pearl-River and Jing-Jin-Ji sites live.

真实世界有多少种双手作业，我们就有多少种数据As many datasets as there are kinds of real two-handed work

工厂只是其中一小部分。我们的采集网络覆盖制造、施工、物流、餐饮、零售与个人服务，并深入医疗手术、兽医、农业户外等小众长尾场景——全部第一人称视角。Factories are only a small part. Our network spans manufacturing, trades, logistics, F&B, retail and personal services, and reaches into surgery, veterinary and outdoor agriculture — all egocentric.

制造装配ManufacturingManufacturing

汽车与设备维修AutomotiveAutomotive & Maintenance

建筑与技术工种ConstructionConstruction & Trades

物流与仓储LogisticsLogistics & Warehousing

餐饮后厨Food & BeverageFood & Beverage

酒店与清洁服务HospitalityHospitality & Cleaning

零售与商业服务RetailRetail & Commercial

家庭维修保养Home RepairHome Repair & Maintenance

农业与户外作业AgricultureAgriculture & Outdoor

手工艺与创作ArtisanalArtisanal Industry

个人护理服务Personal CarePersonal Services

医疗 / 兽医等小众场景Medical / VeterinaryMedical & Niche

NICHE · MEDICAL / HEALTHCARE

连医疗手术与兽医场景，都能采

Even surgery and veterinary scenarios

高精度、强合规要求的医疗作业同样在我们的采集范围内：手术器械打包、外科手术操作、兽医宠物护理等。配合 PII 自动脱敏与企业级数据合规流程，敏感场景也能安全交付。

High-precision, compliance-heavy medical work is within scope: surgical instrument packing, surgical procedures, veterinary pet care and more. Automated PII masking and enterprise compliance make sensitive scenes safe to deliver.

第一人称采集样例（点击播放）Egocentric Capture Samples (click to play)

悬停预览，点击进入全屏播放。全部为真实作业现场第一人称画面。Hover to preview, click for fullscreen. All real work-site, first-person footage.

结构化交付，直连主流 VLA / 模仿学习模型Structured delivery, ready for leading VLA & imitation models

输出标准 MCAP / JSON 与主流数据格式，配套仿真环境，无需再做繁琐的数据清洗与对齐。Standard MCAP / JSON output in mainstream formats, with simulation environments — no tedious re-cleaning or re-alignment.

无论采购整套方案，还是成为采集伙伴Buy the full stack, or join as a capture partner

我们既交付设备与数据管线，也长期通过采集网络收购小众、高价值的第一人称真实作业数据——两条通道都开放。We deliver hardware and data pipelines, and continuously buy niche, high-value egocentric real-task data through our capture network — both channels are open.