Robot training data in China
31 companies with workers here, 31 datasets collected here and 61 sourced events. Reported robot-data pay runs $1.76 to $12.19 an hour, 1.1× to 7.9× the local minimum wage.
Open on the interactive map →Work types
robotics datadata collection audio videoannotation
Minimum wage
- $268 a month, about $1.55 an hour (ILOSTAT, 2024)
Pay reported
| Work | USD per hour | × min wage | Year | Source |
|---|---|---|---|---|
| Day-rate collector | $1.76 to $2.11 | 1.3× | 2026 | sina. |
| Agent-recruited collector, ~10 h/day, 147 yuan net | $2.11 | 1.4× | 2026 | sina. |
| Suqian community home collector (6 h/day) | $2.44 to $2.84 | 1.7× | 2026 | nbd. |
| Home chore video, Shandong | $2.82 | 1.8× | 2026 | restofworld. |
| UMI wearable-sensor data collectors | $2.82 | 1.8× | 2026 | thestandard. |
| Collection-site data collector, base day rate | $3.52 | 2.3× | 2026 | woshipm. |
| Part-time data collector (sorting blocks, stacking cups) | $3.52 | 2.3× | 2026 | qq. |
| Crowdsourced platform task rates (parcel unboxing 30, tea/kitchen 20) | $2.82 to $4.23 | 2.3× | 2026 | qq. |
| Lower-tier-city mothers recording chores | $3.52 to $4.23 | 2.5× | 2026 | sina. |
| Boss Zhipin listings for data collectors | $2.64 to $5.28 | 2.6× | 2026 | qq. |
| Home data collector (cost to middleman) | $4.23 | 2.7× | 2026 | qq. |
| Home collector (mother of two) | $4.33 | 2.8× | 2026 | bjnews. |
| Home/outdoor wearable data collector, Suzhou | $4.40 | 2.8× | 2026 | qq. |
| Staffing-agency direct-hire collector, Beijing | $3.52 to $5.28 | 2.8× | 2026 | sina. |
| Night-shift collector on outsourced contract for an internet major | $6.51 | 4.2× | 2026 | qq. |
| Mocap-glove data collector, Hubei centre | from $8.13 | 5.3× | 2026 | chinadaily. |
| Data collector, industry range | $6.50 to $12.19 | 6.0× | 2026 | cbndata. |
| Part-time home collector | $9.86 | 6.4× | 2026 | sina. |
Companies with workers here
DatatangDataOcean (Haitian Ruisheng)Mecka AIBuild AIAgiBotBeijing Humanoid CenterKylin Training GroundLeju centresHubei state centreJD.comX Square RobotVision LabsManiformerUBTech RoboticsUnitree RoboticsGalbotBeijing Humanoid Robot Innovation Center data & training baseShijingshan Humanoid Robot Data Training CenterHubei Humanoid Robot Innovation CenterQingdao Humanoid Robot Data Collection Training GroundHefei Embodied Robot Data Collection Training GroundJinan Humanoid Robot Data Training CenterWuwen AIPaxini Tianjin super data factoryLightwheelJiyuan ZhihangIO-AI TechGalaxea / Yishu IntelligenceNoitom RoboticsXiaomi RoboticsFourier
Datasets collected here
- AgiBot World (Beta) 1,003,672 trajectories, 2,976.4 hours, 217 tasks huggingface.
co ↗ Major outlet - RoboMIND 107k trajectories, 479 tasks, 96 object classes, 5k failure demos arxiv.
org ↗ Major outlet - Galaxea Open-World Dataset arxiv.
org ↗ Major outlet - HOI4D 2.4M RGB-D frames, 4000 sequences, 800 objects arxiv.
org ↗ Major outlet - Humanoid Robot Kylin Training Ground Target 10 million physical data entries by end 2025 interestingengineering.
com ↗ Major outlet - Beijing Shijingshan humanoid training camp More than 10,000 square meters, 16 scenarios restofworld.
org ↗ Major outlet - Hubei government humanoid data collection centre Hundreds of repetitions a day per task restofworld.
org ↗ Major outlet - China state robot data collection centre network More than 40 centres announced, about two dozen operating restofworld.
org ↗ Major outlet - China UMI wearable data collection facilities 30+ facilities thestandard.
com. Major outlethk ↗ - Liuzhou humanoid training centre nst.
com. Major outletmy ↗ - Vision Labs industrial factory footage backscoop.
com ↗ Secondary - AgiBot World 1M+ trajectories from 100 robots, 100+ scenarios siliconangle.
com ↗ Secondary - EgoSuite-Open100K 100,000 hours; 15,000+ scenes and tasks xinhuanet.
com ↗ Secondary - EgoLive first batch of a 10M-hour programme bjnews.
com. Secondarycn ↗ - OpenLET 60,000+ minutes cjn.
cn ↗ Secondary - HiPHI 617.5 hours (371.8 h whole-body, 245.7 h HOI) eweek.
com ↗ Secondary - G0-Dataset (XRZero-G0) helpnetsecurity.
com ↗ Secondary - Xiaomi UMI pretraining corpus 100,000 hours UMI + ~10,000 h post-training qq.
com ↗ Secondary - MEgo no-embodiment data 1M+ hours ifeng.
com ↗ Secondary - RoboPocket distributed collection 100,000+ hours, 50+ cities ifeng.
com ↗ Secondary - UnifoLM-WBT-Dataset 340 hours, 1.89M trajectories (Mar 2026) thepaper.
cn ↗ Secondary - Fourier ActionNet 30,000+ trajectories (first batch) qbitai.
com ↗ Secondary - Open-H-Embodiment 750 hours; 120,000 trajectories huggingface.
co ↗ Major outlet - Galbot human + feedback data (AstraData) 1M hours human data; 80,000 hours real-world feedback kr-asia.
com ↗ Secondary - UBTech industrial data platform target 10M+ data points per year gasgoo.
com ↗ Secondary - AgiBot World (Colosseo paper) 1M+ trajectories; 217 tasks; five deployment scenarios arxiv.
org ↗ Major outlet - AgiBot World Alpha 100,000+ trajectories; 300 hours; 100 robots huggingface.
co ↗ Major outlet - AgiBot World 2026 huggingface.
co ↗ Major outlet - RoboMIND 2.0 310K+ trajectories; six embodiments; 739 tasks; 12K tactile episodes; 20K mobile; 20K sim arxiv.
org ↗ Major outlet - Unitree G1/Z1 LeRobot datasets about 60 public task datasets on Hugging Face (Feb 2025 to Apr 2026), e.g. Make_The_Bed, Put_Clothes_into_Washing_Machine huggingface.
co ↗ Major outlet - ARIO (All Robots In One) about 3M episodes; 258 series; 321,064 tasks arxiv.
org ↗ Major outlet
Events
- 2026-09-29 106 centralised data centres built in China, 84 operating (STAR Market Daily)
Typical model: local state funds build, robot firms buy back data. qq.com ↗ Secondary - 2026-09-27 Pakistan's World Context records 8,000 factory workers
Wearable-glove and head-camera data sold to US and Chinese robotics firms. arabnews.pk ↗ Secondary - 2026-09-23 AgiBot pledges RM1 billion in Malaysia
Pledged during the Prime Minister's Hangzhou visit, with 5,000 jobs. theedgemalaysia.com ↗ Secondary - 2026-09-20 Reuters: regulators slow humanoid IPOs; data-centre revenue scrutinised
Valuations could fall 60-70% without data-centre revenue. moneycontrol.com ↗ Secondary - 2026-09-10 National Data Administration holds embodied-AI symposium
Pledges embodied-AI data standards work. sina.com. Secondarycn ↗ - 2026-09-09 China recognises 'embodied intelligence robot technician' occupation
MOHRSS adds 11 new occupations. globaltimes.cn ↗ Secondary - 2026-09-09 Mech-Mind CEO alleges data-centre revenue props up IPO candidates
WeChat post triggers scrutiny of state-funded centres. moneycontrol.com ↗ Secondary - 2026-09-03 Beijing X-Humanoid RoboMIND passes 20M downloads
300,000+ dual-arm trajectories over 700+ tasks. people.com. Secondarycn ↗ - 2026-09-01 Reuters: Liuzhou robot training center struggles
Subsidy-dependent center finds a novice trainer needs about 300 attempts per usable movement. nst.com. Major outletmy ↗ - 2026-08-31 Reuters: Liuzhou training centre has no clear path to profitability
Subsidy-dependent centre; robots at 20% of human output. insurancejournal.com ↗ Secondary - 2026-08-29 Shijingshan centre phase 1 rebuilt after original partner exits
Shift to 4D light-field capture and world models. sina.com. Secondarycn ↗ - 2026-08-28 Galbot says it holds 1M hours of human data
Galbot opens simulation, collection hardware and models to partners. kr-asia.com ↗ Secondary - 2026-08-28 Beijing's first humanoid data training centre reported suspended
Shijingshan centre to be rebuilt around new capture methods. sciencenet.cn ↗ Secondary - 2026-08-26 Lightwheel open-sources EgoSuite-Open100K
100,000-hour human behaviour dataset; 10bn-hour five-year plan. xinhuanet.com ↗ Secondary - 2026-08-26 World Humanoid Robot Games full dataset released (2,500+ hours)
Multi-institution training and competition data. youth.cn ↗ Secondary - 2026-08-20 CAICT report: 70+ embodied-AI training grounds operating, 46 more planned
Zhejiang leads with 14; real-robot teleop cost ~275 yuan per effective hour. sina.com. Secondarycn ↗ - 2026-08-19 China recruits UMI data collectors at 20 yuan an hour
Wearable-sensor collectors staff 30+ facilities in Beijing, Shanghai, Guangzhou and other cities. thestandard.com. Major outlethk ↗ - 2026-08-19 Noitom releases 617.5-hour HiPHI motion dataset at WRC
132 performers, 90 Hz, free release. eweek.com ↗ Secondary - 2026-08-19 Unitree soars 460% on Shanghai STAR debut
IPO at $9B valuation; ~$66B after first day. yahoo.com ↗ Secondary - 2026-08-17 Maniformer raises several hundred million yuan led by China Telecom
Funds MEgo no-embodiment capture scale-up. qq.com ↗ Secondary - 2026-08-11 OpenLET national open embodied dataset community opens Wuhan working group
Leju-led community shares 60,000+ minutes of full-size humanoid data. cjn.cn ↗ Secondary - 2026-08-05 MIIT approves first embodied-AI dataset quality standard
Industry standard effective 1 Nov 2026, drafted with 40+ organisations. xinhuanet.com ↗ Secondary - 2026-08-05 Guizhou policy subsidises opening scenes for embodied data collection
50,000 yuan per opened scene, 500,000 yuan for demonstration scenes. sina.com. Secondarycn ↗ - 2026-07-22 China starts first national standard on humanoid dataset quality
TC591 project (dataset part 6: quality evaluation) led with UBTech. sina.com. Secondarycn ↗ - 2026-07-16 Xiaomi model pretrained on 100,000 hours of UMI data
Xiaomi-Robotics-1 uses robot-free UMI data plus ~10,000 h cross-embodiment data. qq.com ↗ Secondary - 2026-07-09 Forecast of 1 million new embodied-AI jobs including data collectors
Over five years, per report in 21st Century Business Herald. 21jingji.com ↗ Secondary - 2026-06-30 Shandong lists first 26 provincial embodied-AI training grounds
Seven Jinan projects selected. qq.com ↗ Secondary - 2026-06-26 Korean data firms shift labelling to Vietnam
Driving and robotics labelling moves to lower-cost countries. ajupress.com ↗ Secondary - 2026-06-25 AgiBot releases 'Data Collection 2.0' at Chengdu Pidu centre
Centre opened May 2026, supports 430+ scenes. scol.com. Secondarycn ↗ - 2026-06-25 BitRobot, Hugging Face and Unitree release HIW-500 in-the-wild humanoid dataset
500+ hours of G1 whole-body teleoperation in real Southeast Asian homes. zhidx.com ↗ Secondary - 2026-06-17 WIRED visits IO-AI Tech teleoperation floor in Shenzhen
Workers control humanoids with VR and body tracking. wired.com ↗ Major outlet - 2026-06-16 Galaxea and Beijing E-Town launch 1M-hour real data plan
Joint data company Yishu Intelligence formed. jjckb.cn ↗ Secondary - 2026-06-09 MIIT and SASAC launch 2026 real-scene training special action
Each province must pick at least 20 scene units; central SOEs at least 10. cnfin.com ↗ Secondary - 2026-06-09 MIIT and SASAC special action plan on real-world training of humanoids
Targets capacity to deploy 10,000 units by end-2026. biggo.com ↗ Secondary - 2026-06-03 JD.com and Suqian government plan 10 million hours of robot data
Two-year target for robotics training data in founder's hometown. restofworld.org ↗ Major outlet - 2026-05-20 JD.com opens China's first embodied-AI data collection community in Suqian
Residents record chores for 3,000-3,500 yuan a month. nbd.com. Secondarycn ↗ - 2026-05-20 Chinese report: robot makers sell to state data centres, then buy the data back
64 data collection centres in 27+ cities by April 2026; subsidies cover robot firms' data purchases. qq.com ↗ Secondary - 2026-03-27 Unitree open-sources UnifoLM-WBT whole-body teleop dataset
Full-body teleoperation data from G1 in homes and factories, Apache-2.0 on Hugging Face. thepaper.cn ↗ Secondary - 2026-03-27 Unitree G1/Z1 LeRobot datasets released
Unitree G1/Z1 LeRobot datasets: about 60 public task datasets on Hugging Face (Feb 2025 to Apr 2026), e.g. Make_The_Bed, Put_Clothes_into_Washing_Machine huggingface.co ↗ Major outlet - 2026-03-20 China's robot data center network expands
Training hubs open across Anhui, Zhejiang, Shandong, Hebei and Hubei. humanoid.guide ↗ Secondary - 2026-03-11 AgiBot World 2026 released
AgiBot World 2026: free-form real-world robot data with digital twins huggingface.co ↗ Major outlet - 2026-02-24 Xinhua profiles Qingdao data collection training ground
40+ collectors, target over 1 million real-robot records a year. xinhuanet.com ↗ Secondary - 2026-02-06 Open-H-Embodiment released
Open-H-Embodiment: 750 hours; 120,000 trajectories; 4.5 TB huggingface.co ↗ Major outlet - 2026-01-10 Qingdao humanoid data training ground in trial operation
1,500 m2 site with 28 stations opened in early Dec 2025. qingdaonews.com ↗ Secondary - 2026-01-07 Rest of World: China's state-funded robot training centers
Over 40 state-owned robot data centers announced; trainers repeat motions in VR headsets and exoskeletons. restofworld.org ↗ Major outlet - 2026-01-07 Rest of World: 40+ state-owned robot data collection centres announced in China
Interact Analysis counted 40+ announced centres by Dec 2025, ~two dozen operating. restofworld.org ↗ Major outlet - 2025-12-31 RoboMIND 2.0 released
RoboMIND 2.0: 310K+ trajectories; six embodiments; 739 tasks; 12K tactile episodes; 20K mobile; 20K sim arxiv.org ↗ Major outlet - 2025-12-04 Chinese AI firms hire Kenyan video labellers informally
Kenyan students label short videos for China-based firms for ~KES 700 a day. restofworld.org ↗ Major outlet - 2025-11-17 UBTech Walker S2 orders top 800M yuan, incl. 159M yuan Zigong data centre
Second-largest 2025 order was humanoids for a data collection center in Zigong. prnewswire.co. Secondaryuk ↗ - 2025-10-09 Beijing opens 10,000 sq m humanoid training center
Shijingshan facility with 16 replica scenarios claims 6 million data points a year. beijing.gov. Major outletcn ↗ - 2025-08-23 Galaxea Open-World Dataset released
Galaxea Open-World Dataset: 500+ hours; 100K trajectories; 150 task categories; 1,600+ objects huggingface.co ↗ Major outlet - 2025-04-09 Beijing's first humanoid robot data training centre opens at Shougang Park
About 100 robots, each collecting about 100 records a day. beijing.gov. Major outletcn ↗ - 2025-03-19 National Data Administration reports 7 annotation bases
Bases in Chengdu, Shenyang, Hefei, Changsha, Haikou, Baoding and Datong employ 58,000 annotators. nda.gov. Major outletcn ↗ - 2025-03-17 Fourier open-sources ActionNet humanoid dataset
First batch of 30,000+ real-robot trajectories plus full collection-to-deployment toolchain. qbitai.com ↗ Secondary - 2025-03-09 AgiBot World (Colosseo paper) released
AgiBot World (Colosseo paper): 1M+ trajectories; 217 tasks; five deployment scenarios arxiv.org ↗ Major outlet - 2025-03-08 Fourier ActionNet released
Fourier ActionNet: 30K+ trajectories; about 140 hours action-net.org ↗ Secondary - 2025-02-11 AgiBot World Beta released
AgiBot World Beta: 1M+ trajectories; 2976.4 hours; 100 robots huggingface.co ↗ Major outlet - 2024-12-30 AgiBot World dataset alpha released
Data from AgiBot's replicated-scene collection setup, released as alpha in Dec 2024 and beta (1M+ trajectories) on 1 March 2025. github.com ↗ Major outlet - 2024-12-30 AgiBot World Alpha released
AgiBot World Alpha: 100,000+ trajectories; 300 hours; 100 robots huggingface.co ↗ Major outlet - 2024-12-18 RoboMIND released
RoboMIND: 107k trajectories; 479 tasks; 96 object classes; 5k failure demos arxiv.org ↗ Major outlet - 2024-08-20 ARIO (All Robots In One) released
ARIO (All Robots In One): about 3M episodes; 258 series; 321,064 tasks arxiv.org ↗ Major outlet
On-the-ground evidence
- AgiBot (Zhiyuan), Shanghai data collection site
Operating 17 hours a day, the site's goal is to generate reams of data that its owner, Chinese humanoid startup AgiBot, uses to train robots
reuters.com ↗ Major outlet - Beijing Humanoid Robot Innovation Center data collection base
北京人形机器人创新中心数采基地占地"近5000平方米"，部署"机器人120多台"
woshipm.com ↗ Secondary - Beijing Shougang Park humanoid robot data training centre (RealMan robots)
每个机器人平均每天会采集大概100条数据
beijing.gov. Major outletcn ↗ - Hubei humanoid robot innovation center (Optics Valley, Wuhan)
wearing motion-capture gloves to teach robots how to grasp objects and walk
chinadaily.com. Major outletcn ↗ - Qingdao humanoid robot data collection training ground (Laoshan)
青岛市人形机器人数据采集训练场有40多位数据采集员
xinhuanet.com ↗ Secondary - Sichuan Humanoid Robot Multimodal Data Collection and Testing Center (Zigong), UBTech Walker S2
Training a robot to sort a single battery requires tens of thousands of grasping actions.
theborneopost.com ↗ Secondary - Hefei embodied AI robot data collection pre-training ground
合肥市具身智能机器人数据采集预训练场
news.cn ↗ Major outlet - JD.com (Suqian data programme)
JD.com, which operates an Amazon-style logistics network, said it plans to eventually bring 100,000 employees and 500,000 external workers to join its data collection operations.
restofworld.org ↗ Major outlet - Guangdong (Dongguan) data vendors working with factories
Two data vendors in the southern province of Guangdong, a manufacturing hub, told Rest of World they were working with dozens of electronics and packaging factories to have workers wear data-collection devices.
restofworld.org ↗ Major outlet - IO-AI Tech (Shenzhen)
Employees don headsets reminiscent of Ready Player One. They control robots in mock convenience stores and factory lines.
webpronews.com ↗ Secondary - MiFeng Tech (MEgo egocentric devices)
第2万套MEgo无本体数采设备下线
sciencenet.cn ↗ Secondary - Appen (China its biggest robotics-data market)
China was Appen's biggest market for robotics data with the US second
abc.net. Major outletau ↗
Every row links to its source, and hovering a link shows the quote it rests on. Research compiled 2026-10-05. Method