What the data shows
- 1
Tesla has the most documented in-house data operation (13 sourced facts).
- 2
16 companies and centres run a dedicated data-collection facility on record (18 facility facts).
- 3
14 published pay figures for in-house collectors and operators from 11 companies.
- 4
23 stated team sizes and data volumes, from small teleoperation teams to datasets measured in hundreds of thousands of hours.
Figure 1
Who collects their own robot data, and how
Which robotics companies run their own data collection, with what method, how many people and how many hours?
| Company | Method | Sourced facts |
|---|---|---|
| TeslaAUTO BILD | Tesla told staff at Gigafactory Berlin-Brandenburg (Gruenheide) that selected teams would wear cameras during their work so the footage can train Optimus, per Handelsblatt's report of a company email. | 13 |
| Figure AIFigure AI (Greenhouse job board) | Figure runs offsite robot data collection sessions across an expanding portfolio of residential sites in the Bay Area, coordinated on a night shift. | 10 |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)The Standard (Hong Kong) | Each robot at the Shijingshan center is paired with a dedicated trainer who teleoperates it with VR headset and controllers while joint angles and grip force are recorded; methods also include exoskeleton teleoperation and first-person capture. | 8 |
| AgiBotYicai Global | AgiBot's data subsidiary Maniformer launched MEgo collection hardware (grippers and head-mounted devices) so human operators wearing sensors can record tasks in real environments without a physical robot, targeting tens of millions of hours of data in 2026. | 6 |
| Physical IntelligenceAshby (Physical Intelligence careers) | Physical Intelligence Robot Operators either lead the robot with their own arms (the robot mirrors them) or use specialized controllers, and the job is metrics-based with data collection goals per shift. | 6 |
| Sunday RoboticsSunday Robotics | Sunday Robotics trains its Memo robot on data from Memory Developers wearing its Skill Capture Glove at home, and says it has shipped thousands of gloves to contributors across the country. | 5 |
| 1X Technologies1X Technologies | For tasks NEO does not know, a 1X Expert can remotely supervise the robot in the owner's home at scheduled times, which 1X says helps NEO learn new abilities. | 4 |
| Generalist AIGeneralist AI | Generalist AI said its robot data operations add over 10,000 new hours per week, powered by a global network of thousands of data collection devices and robots. | 4 |
| OpenAIBusiness Insider | OpenAI's data collectors use 3D-printed GELLO controllers to teleoperate two Franka robot arms, and are rated on how many good hours of functional training data they generate, Business Insider reported. | 4 |
| UberTechCrunch | Uber said it plans to roll out 500 data-collection Hyundai Ioniq 5 vehicles globally in 2026 (50 on the road by summer), each fitted with 14 cameras, eight solid-state lidars and nine radars, to gather training data for its autonomous vehicle partners. | 4 |
| Agility RoboticsAgility Robotics (Greenhouse job board) | Agility mixes internal and contract labor for data collection and weighs external data collection vendors against in-house capacity on cost and quality (build versus buy). | 3 |
| ApptronikApptronik | Apptronik's Apollo 2 robots generate training data at Robot Park through a mix of teleoperation and autonomous execution; the data trains Google DeepMind's Gemini Robotics models. | 3 |
| Dyna RoboticsAshby (Dyna Robotics careers) | Dyna Robotics data collectors manually control robot arms (teleoperation) across a diverse range of tasks and make sure the recorded data is annotated correctly. | 3 |
| Google DeepMindGoogle DeepMind | Google DeepMind's AutoRT system ran up to 20 robots at once and 52 robots in total across office buildings for seven months, gathering 77,000 robot trials across 6,650 tasks through a mix of teleoperation and autonomous policies. | 3 |
| Hubei Humanoid Robot Innovation CenterChina Daily (Xinhua) | At the Hubei center, trainers wear VR headsets to remote-control the robots, and the data collection section has mock kitchens, bathrooms, living rooms, a milk tea shop, a pharmacy, a restaurant and a warehouse. | 3 |
| LG ElectronicsThe Investor (Korea Herald) | LG's Yangjae data factory in Seoul, due for full operation by end 2026, has CLOiD robots collecting data in a mock home and in a replica of LG's Tennessee washing machine plant, plus spaces for LG CNS logistics and LG Innotek robotic hands. | 3 |
| Liuzhou humanoid robot training centre (Guangxi; UBTech hardware)New Straits Times (Reuters) | At the Liuzhou centre a novice trainer gets about one usable movement in 300 attempts and an experienced trainer about one in 50; trainers see through the robot's eyes via headsets and drive it with sensor-fitted handheld controllers. | 3 |
| Toyota Research InstituteToyota USA Newsroom | Toyota Research Institute said its robots learn new skills from haptic demonstrations by a human teacher using Diffusion Policy, and that it had taught over 60 skills with a target of 1,000 by the end of 2024. | 3 |
| WaymoCNBC (via NBC New York) | Waymo confirmed it would test in Philadelphia through fall 2025 with a limited fleet and human safety drivers, first driving manually through downtown and freeways; CNBC reports the road trips collect mapping data and test the Waymo Driver in new environments. | 3 |
| ZiplineZipline (Greenhouse job board) | Zipline runs a flight test facility in Amarillo, Texas staffed by Flight Test Operator, Ground Test and Visual Observer teams, where aircraft move from engineering development into flight test; it is hiring a Site Lead to run it. | 3 |
| Beijing Innovation Center of Humanoid RoboticsGlobal Times | The Beijing Innovation Center of Humanoid Robotics runs a training facility of nearly 6,000 square metres with more than 30 scenarios and more than 150 robot units in 40 configurations for home, retail, industrial and medical data collection. | 2 |
| Chinese state-backed robot data collection centresRest of World | By December 2025 more than 40 state-owned robot data collection centres had been announced in China, about two dozen of them operating, typically built by local governments and run by robotics companies, per Interact Analysis. | 2 |
| National and Local Co-Built Humanoid Robotics Innovation Center (Shanghai)Pudong New Area government (China Daily) | Shanghai's first heterogeneous humanoid robot training ground (formerly the Kylin Training Ground), run under the National and Local Co-Built Humanoid Robotics Innovation Center, opened in January 2025 on 4,600 square metres with more than 100 robots in 10 scenarios. | 2 |
| Skild AISkild AI | Skild AI said it is scaling robot teleoperation, UMI handheld grippers, egocentric video and simulation data collection in-house, because no single source wins on scalability, diversity and closeness to the robot. | 2 |
| SkydioSkydio (Ashby job board) | Skydio is hiring a fixed-term (12-month) Data Collection Flight Test Operator in San Mateo, California to fly its drones across varied US environments, about 90% of the time on the road, gathering the data that trains and validates its autonomy and AI. | 2 |
| WaabiWaabi (Lever job board) | Waabi's Vehicle Operator (CDL) day-shift role in Dallas, Texas, which manually drives Class 8 trucks on mapping and shadow missions and supervises the Waabi Driver, pays $31 to $43 an hour. | 2 |
| WayveWayve (Ashby job board) | Wayve is hiring Vehicle Safety Operators in Tokyo to monitor its AI driving system on public roads, intervene when required and support data collection that trains its AI. | 2 |
| Boston DynamicsBoston Dynamics | Boston Dynamics collects Atlas training data in-house with its own VR teleoperation system, combining Atlas' model predictive controller with a custom VR interface for tasks from fine finger work to whole-body reaching and walking. | 1 |
| GalbotChina Daily | Galbot's co-founder said the company pre-trains on large synthetic datasets and fine-tunes with minimal real-world data, because real-world data collection is prohibitively expensive. | 1 |
| Kodiak AIKodiak (Greenhouse job board) | Kodiak's Class A Safety Driver role at its Lancaster, Texas operations pays $37 an hour on a five days on, two days off schedule, with a $3,000 referral bonus per hired driver. | 1 |
| MotionalMotional (Greenhouse job board) | Motional is hiring Autonomous Driving Test Operators for its growing AV Test Operations team (posting listed in Pittsburgh) to operate autonomous vehicles, collect and annotate data and give engineers feedback; the posted base salary is $52,000. | 1 |
| NuroNuro (Greenhouse job board) | Nuro's Autonomous Vehicle Operator role in South San Francisco (12-month temporary assignment) pays $27 an hour on morning/afternoon shifts and $28.25 overnight; operators drive test vehicles in manual and autonomous modes up to eight hours a day and collect real-world driving data. | 1 |
| RealMan RoboticsElectropages | A Beijing Humanoid Robotics Data Training Centre where RealMan Robotics is the core equipment provider and operator opened on 3,000 square metres with 108 robots and ten real-world scenes, estimated to produce over one million data points a year. | 1 |
| ZooxAutomotive World | Zoox said it would start testing in Dallas and Phoenix with retrofitted Toyota Highlander SUVs and safety drivers that map roads and collect operational data, taking its US testing footprint to ten cities. | 1 |
Base: 34 companies and centres, 115 facts, each linked to its source (reviewed 18 Oct 2026).
A company appears once a source documents data it collects with its own people; vendors that collect for others are in the vendors report.
Credit: compiled by RobotTraining.jobs Research, How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay, checked 18 Oct 2026.
Figure 2
How much: teams and hours of data
How many people do robotics companies put on data collection, and how many hours of data have they gathered?
| Company | Measure | As stated |
|---|---|---|
| 1X Technologies1X Technologies | Hours collected | 1X said it had gathered thousands of hours of data over a year on its EVE humanoids doing mobile manipulation tasks in homes and offices. |
| AgiBotarXiv | Hours collected | AgiBot World, collected by more than 100 identical AgiBot G1 robots at AgiBot's own facility, contains 1,001,552 trajectories totalling 2,976.4 hours across 217 tasks. |
| AgiBotKrASIA (36Kr English) | Team size | AgiBot assigns one human instructor, a skilled data collector (typically a young professional with strong physical coordination), to each robot at its Shanghai data acquisition center. |
| Beijing Innovation Center of Humanoid RoboticsGlobal Times | Hours collected | The center's training base had delivered nearly 30,000 hours of robot data to external partners, with more than 95 percent passing quality checks, and is working toward a goal of 1 million hours. |
| Figure AIFigure AI | Hours collected | Figure's Helix 02 whole-body controller (System 0) was trained on over 1,000 hours of human motion data plus sim-to-real reinforcement learning. |
| Figure AIFigure AI | Hours collected | Figure trained its Helix VLA model on a multi-robot, multi-operator dataset of teleoperated behaviors totalling about 500 hours. |
| Generalist AIGeneralist AI | Hours collected | By the GEN-1 release, Generalist AI's dataset had grown to over half a million hours of physical interaction data. |
| Generalist AIGeneralist AI | Hours collected | Generalist AI said GEN-0 was pretrained on its in-house dataset of over 270,000 hours of real-world manipulation data, collected in thousands of homes, warehouses and workplaces, partly together with data foundry partners. |
| Google DeepMindarXiv (Google DeepMind) | Hours collected | Google DeepMind collected a teleoperated robot action dataset on a fleet of ALOHA 2 robots over 12 months, thousands of hours of expert demonstrations, to train Gemini Robotics. |
| Hubei Humanoid Robot Innovation CenterAcademic Divisions of the Chinese Academy of Sciences | Hours collected | The Hubei center says it can collect over 1 million real-machine data entries a year, which are verified, labelled and cleaned before model training; a July 2026 visit report put output at 24,000 valid data records per day. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)Global Times | Hours collected | At the Shijingshan center an eight-hour shift first yielded about 0.9 hours of usable data per data collector; after process upgrades the yield rose to five to six hours a day, per the center's figures cited by an analyst. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Hours collected | The Shijingshan center can produce more than 8 million multimodal data records a year, and Leju says it produces nearly 25 million real-machine data records a year across the facilities it helped build. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Team size | The Shijingshan humanoid data-training center co-built by Leju Robotics and the Shijingshan district government had an instructor team of more than 200 people, still expanding, per Leju's training-center technical director. |
| LG ElectronicsThe Investor (Korea Herald) | Hours collected | LG Electronics expects to collect 100,000 hours of robot training data with its own CLOiD robots at its Yangjae data factory in Seoul by the end of 2026. |
| OpenAIBusiness Insider | Team size | Business Insider reported that OpenAI's robotics lab in San Francisco employs around 100 data collectors and has more than quadrupled since it launched in February 2025. |
| Physical IntelligencearXiv (Physical Intelligence) | Hours collected | For pi0.5, Physical Intelligence used about 400 hours of data from its mobile manipulators doing household tasks in about 100 different homes. |
| Physical IntelligencearXiv (Physical Intelligence) | Hours collected | Physical Intelligence pre-trained its pi0 model on about 10,000 hours of dexterous manipulation data from its own robots (7 robot configurations, 68 tasks), on top of open datasets such as OXE, DROID and Bridge. |
| Sunday RoboticsSunday Robotics | Team size | Sunday Robotics said that ahead of its beta program it had grown Engineering 3x, Research 4x and Data Operations 5x. |
| TeslaBusiness Insider | Hours collected | Each Optimus data collector is expected to gather at least four hours of usable video footage per shift, five current and former workers said. |
| TeslaBusiness Insider | Team size | Tesla registered 224 in-house test drivers and 104 vehicles with the California DMV in December 2024 for its autonomous testing permit, up from 59 drivers and 14 vehicles, according to public records viewed by Business Insider. |
| TeslaBusiness Insider | Team size | More than 50 different people worked as Optimus data collection operators at Tesla over the year to August 2024, per a Business Insider review of LinkedIn data. |
| Toyota Research InstitutearXiv (Toyota Research Institute) | Hours collected | Toyota Research Institute trained its Large Behavior Models on about 1,700 hours of demonstrations, of which about 545 hours (532 tasks, 64,262 demonstrations) were collected in-house at TRI. |
| UberTechCrunch | Hours collected | Uber said its planned 500-vehicle AV Labs fleet will be able to collect 2 million miles of high-fidelity driving data per month (a stated capacity, not miles already collected). |
Base: 23 team sizes and data volumes, each in its source's own words.
Volumes are as each company states them (hours of video, hours of teleoperation, episodes); they are not converted to one unit.
Credit: compiled by RobotTraining.jobs Research, How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay, checked 18 Oct 2026.
Figure 3
What in-house data collectors are paid
What do robotics companies pay the people who collect their training data in-house?
| Company | Pay as stated |
|---|---|
| 1X Technologies1X Technologies (Ashby job board) | 1X's Operator, Data Collection (Swing Shift) role in San Carlos, on the Data Collection Operations team that operates NEO robots in VR, pays $28 to $36 per hour plus equity. |
| Dyna RoboticsAshby (Dyna Robotics careers) | Dyna Robotics advertises an on-site Data Collection Operator role at its Redwood City headquarters, teleoperating robot arms to produce training data, paying $30 to $32 an hour. |
| Figure AIFigure AI (Greenhouse job board) | Figure's Humanoid Robot Pilot role in San Jose (wear teleoperation equipment, guide the robot through behaviors, upload the data for AI training) is a 6-month fixed-term hourly job at a US base pay of $32 per hour. |
| Kodiak AIKodiak (Greenhouse job board) | Kodiak's Class A Safety Driver role at its Lancaster, Texas operations pays $37 an hour on a five days on, two days off schedule, with a $3,000 referral bonus per hired driver. |
| NuroNuro (Greenhouse job board) | Nuro's Autonomous Vehicle Operator role in South San Francisco (12-month temporary assignment) pays $27 an hour on morning/afternoon shifts and $28.25 overnight; operators drive test vehicles in manual and autonomous modes up to eight hours a day and collect real-world driving data. |
| Physical IntelligenceAshby (Physical Intelligence careers) | Physical Intelligence advertises an on-site Robot Operator role in San Francisco that teleoperates robot arms to collect training data, paying $25 an hour on 8-hour evening or overnight shifts, minimum 5 days a week. |
| SkydioSkydio (Ashby job board) | Skydio's Data Collection Flight Test Operator role pays $50 to $60 an hour. |
| Sunday RoboticsAshby (Sunday Robotics careers) | Sunday Robotics pays remote Memory Developers, who record household chores with its hardware, $30 per hour of approved data during a one-week screening and up to $60 per hour for complex tasks afterwards. |
| Sunday RoboticsAshby (Sunday Robotics careers) | Sunday Robotics also hires in-house Memory Developers at Redwood City at $35 an hour on a fixed 5-day, 8-hour evening shift (2 PM to 10 PM), who also pilot new tasks and hardware before they go to the remote collectors. |
| TeslaAUTO BILD | Auto Bild reported that Tesla's US Optimus Data Collection Operators are paid hourly wages between 25.25 and 34.50 dollars. |
| TeslaBusiness Insider (via AOL) | Tesla's 'Vehicle Operator, Autopilot' listing in Queens, New York, for operators who drive engineering cars and collect audio and camera data, paid Test Operator I $25.25 to $27.60 an hour base and Test Operator II $28.75 to $30.60, with 10% more on afternoon and night shifts (up to $33.66). |
| TeslaBusiness Insider | Tesla's Optimus 'Data Collection Operator' job listing offered full-time pay of $25.25 to $48 per hour. |
| WaabiWaabi (Lever job board) | Waabi's Vehicle Operator (CDL) day-shift role in Dallas, Texas, which manually drives Class 8 trucks on mapping and shadow missions and supervises the Waabi Driver, pays $31 to $43 an hour. |
| ZiplineZipline (Greenhouse job board) | Zipline's Flight Test Operator role at its Esparto, California test site, acting as Remote Pilot in Command for high-volume flight testing including nights and weekends, starts at $34 to $38 an hour. |
Base: 14 pay figures from job postings and reporting.
A posting's range is the employer's offer, not measured earnings.
Credit: compiled by RobotTraining.jobs Research, How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay, checked 18 Oct 2026.
Figure 4
Every sourced fact
What has each company or its press coverage said about collecting its own robot data?
| Company | Kind | Fact |
|---|---|---|
| 1X Technologies1X Technologies | Method | For tasks NEO does not know, a 1X Expert can remotely supervise the robot in the owner's home at scheduled times, which 1X says helps NEO learn new abilities. |
| 1X Technologies1X Technologies (Ashby job board) | Pay | 1X's Operator, Data Collection (Swing Shift) role in San Carlos, on the Data Collection Operations team that operates NEO robots in VR, pays $28 to $36 per hour plus equity. |
| 1X Technologies1X Technologies | Method | 1X trains its Redwood model on teleoperated and autonomous episodes gathered on NEO robots in 1X offices and in employees' homes. |
| 1X Technologies1X Technologies | Hours collected | 1X said it had gathered thousands of hours of data over a year on its EVE humanoids doing mobile manipulation tasks in homes and offices. |
| AgiBotYicai Global | Method | AgiBot's data subsidiary Maniformer launched MEgo collection hardware (grippers and head-mounted devices) so human operators wearing sensors can record tasks in real environments without a physical robot, targeting tens of millions of hours of data in 2026. |
| AgiBotarXiv | Hours collected | AgiBot World, collected by more than 100 identical AgiBot G1 robots at AgiBot's own facility, contains 1,001,552 trajectories totalling 2,976.4 hours across 217 tasks. |
| AgiBotarXiv | Method | AgiBot collects its demonstrations with skilled teleoperators using two teleoperation systems, VR headset control and whole-body motion capture control, and annotators then check each episode. |
| AgiBotKrASIA (36Kr English) | Facility | AgiBot runs its own 3,000 square metre data acquisition center in Pudong, Shanghai, with mock bedroom, kitchen, supermarket checkout and other rooms where staff guide humanoid robots through everyday tasks; it planned to add 1,000 square metres. |
| AgiBotKrASIA (36Kr English) | Why | AgiBot built its own data center after finding open-source datasets unsuitable: data from robots with different designs and sensors could not be pooled, so it decided it needed a proprietary source. |
| AgiBotKrASIA (36Kr English) | Team size | AgiBot assigns one human instructor, a skilled data collector (typically a young professional with strong physical coordination), to each robot at its Shanghai data acquisition center. |
| Agility RoboticsAgility Robotics (Greenhouse job board) | Hiring drive | Agility Robotics is hiring a Data Collection Manager to build and scale its in-house teleoperation data collection team of robot operators and data collection specialists in Fremont. |
| Agility RoboticsAgility Robotics (Greenhouse job board) | Method | Agility mixes internal and contract labor for data collection and weighs external data collection vendors against in-house capacity on cost and quality (build versus buy). |
| Agility RoboticsAgility Robotics (Greenhouse job board) | Why | Agility says data collection is the input to its learned policies and that the quality of the operation directly gates model progress. |
| ApptronikApptronik | Facility | Apptronik opened an expanded Robot Park in Austin, its flagship humanoid data collection and training facility of nearly 90,000 square feet, anchoring a network of Robot Parks at partner and customer sites. |
| ApptronikApptronik | Method | Apptronik's Apollo 2 robots generate training data at Robot Park through a mix of teleoperation and autonomous execution; the data trains Google DeepMind's Gemini Robotics models. |
| ApptronikApptronik | Why | Apptronik says humanoids need large amounts of real-world data to train the embodied AI models that let them work autonomously, and Robot Park is where much of it is created. |
| Beijing Innovation Center of Humanoid RoboticsGlobal Times | Facility | The Beijing Innovation Center of Humanoid Robotics runs a training facility of nearly 6,000 square metres with more than 30 scenarios and more than 150 robot units in 40 configurations for home, retail, industrial and medical data collection. |
| Beijing Innovation Center of Humanoid RoboticsGlobal Times | Hours collected | The center's training base had delivered nearly 30,000 hours of robot data to external partners, with more than 95 percent passing quality checks, and is working toward a goal of 1 million hours. |
| Boston DynamicsBoston Dynamics | Method | Boston Dynamics collects Atlas training data in-house with its own VR teleoperation system, combining Atlas' model predictive controller with a custom VR interface for tasks from fine finger work to whole-body reaching and walking. |
| Chinese state-backed robot data collection centresGlobal Times | Why | Chinese public training centres pool equipment, scenarios and personnel because robotics firms that bought equipment, built scenes and collected data in-house found it expensive and inefficient. |
| Chinese state-backed robot data collection centresRest of World | Facility | By December 2025 more than 40 state-owned robot data collection centres had been announced in China, about two dozen of them operating, typically built by local governments and run by robotics companies, per Interact Analysis. |
| Dyna RoboticsAshby (Dyna Robotics careers) | Method | Dyna Robotics data collectors manually control robot arms (teleoperation) across a diverse range of tasks and make sure the recorded data is annotated correctly. |
| Dyna RoboticsAshby (Dyna Robotics careers) | Pay | Dyna Robotics advertises an on-site Data Collection Operator role at its Redwood City headquarters, teleoperating robot arms to produce training data, paying $30 to $32 an hour. |
| Dyna RoboticsDyna Robotics | Why | Dyna Robotics calls robot data the scarcest input because it costs robot and operator hours and loses value with each hardware revision, so its Dyna-2.1 system also learns from human recordings. |
| Figure AIFigure AI (Greenhouse job board) | Pay | Figure's Humanoid Robot Pilot role in San Jose (wear teleoperation equipment, guide the robot through behaviors, upload the data for AI training) is a 6-month fixed-term hourly job at a US base pay of $32 per hour. |
| Figure AIFigure AI (Greenhouse job board) | Method | Figure runs offsite robot data collection sessions across an expanding portfolio of residential sites in the Bay Area, coordinated on a night shift. |
| Figure AIFigure AI (Greenhouse job board) | Facility | Figure is establishing its first data operations unit in Mexico City. |
| Figure AIFigure AI (Greenhouse job board) | Method | Figure's Mexico City data unit captures human data with sensor suits, headsets, cameras and rigs at capture stations maintained on site. |
| Figure AIFigure AI (Greenhouse job board) | Facility | Figure is establishing its first data operations unit in Brazil, in Sao Paulo, staffed by 'Data Creators' who capture training data for its humanoids. |
| Figure AIFigure AI (Greenhouse job board) | Method | Figure's Sao Paulo data unit runs 24/7 in three rotating shifts (6:00 to 15:00, 14:30 to 23:30 and 23:00 to 8:00), with shift coordinators sequencing capture work across Data Creators and stations. |
| Figure AIFigure AI | Hours collected | Figure's Helix 02 whole-body controller (System 0) was trained on over 1,000 hours of human motion data plus sim-to-real reinforcement learning. |
| Figure AIFigure AI | Method | Under Project Go-Big, Figure collects egocentric human video passively as people go about real Brookfield homes, and trained Helix navigation on that video alone. |
| Figure AIFigure AI | Why | Figure says humanoids can learn directly from everyday human video because their perspective and kinematics mirror a person's. |
| Figure AIFigure AI | Hours collected | Figure trained its Helix VLA model on a multi-robot, multi-operator dataset of teleoperated behaviors totalling about 500 hours. |
| GalbotChina Daily | Why | Galbot's co-founder said the company pre-trains on large synthetic datasets and fine-tunes with minimal real-world data, because real-world data collection is prohibitively expensive. |
| Generalist AIGeneralist AI | Hours collected | By the GEN-1 release, Generalist AI's dataset had grown to over half a million hours of physical interaction data. |
| Generalist AIGeneralist AI | Why | Generalist AI said its base model is trained without robot data, on low-cost wearable devices worn by humans, and argued teleoperation produces slower, less fluid data. |
| Generalist AIGeneralist AI | Hours collected | Generalist AI said GEN-0 was pretrained on its in-house dataset of over 270,000 hours of real-world manipulation data, collected in thousands of homes, warehouses and workplaces, partly together with data foundry partners. |
| Generalist AIGeneralist AI | Method | Generalist AI said its robot data operations add over 10,000 new hours per week, powered by a global network of thousands of data collection devices and robots. |
| Google DeepMindarXiv (Google DeepMind) | Hours collected | Google DeepMind collected a teleoperated robot action dataset on a fleet of ALOHA 2 robots over 12 months, thousands of hours of expert demonstrations, to train Gemini Robotics. |
| Google DeepMindarXiv (Google DeepMind) | Why | Google DeepMind researchers wrote that teleoperated data is more diverse and valuable than autonomously collected data but is limited by how many human operators are available, which is why AutoRT mixes teleoperation with autonomous collection. |
| Google DeepMindGoogle DeepMind | Method | Google DeepMind's AutoRT system ran up to 20 robots at once and 52 robots in total across office buildings for seven months, gathering 77,000 robot trials across 6,650 tasks through a mix of teleoperation and autonomous policies. |
| Hubei Humanoid Robot Innovation CenterChina Daily (Xinhua) | Method | At the Hubei center, trainers wear VR headsets to remote-control the robots, and the data collection section has mock kitchens, bathrooms, living rooms, a milk tea shop, a pharmacy, a restaurant and a warehouse. |
| Hubei Humanoid Robot Innovation CenterAcademic Divisions of the Chinese Academy of Sciences | Facility | The Hubei Humanoid Robot Innovation Center in Wuhan, opened June 2025, offers 23 simulation scenarios and more than 10 temporary ones so that over 100 humanoid robots can train at once under data collection specialists. |
| Hubei Humanoid Robot Innovation CenterAcademic Divisions of the Chinese Academy of Sciences | Hours collected | The Hubei center says it can collect over 1 million real-machine data entries a year, which are verified, labelled and cleaned before model training; a July 2026 visit report put output at 24,000 valid data records per day. |
| Kodiak AIKodiak (Greenhouse job board) | Pay | Kodiak's Class A Safety Driver role at its Lancaster, Texas operations pays $37 an hour on a five days on, two days off schedule, with a $3,000 referral bonus per hired driver. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)The Standard (Hong Kong) | Method | Each robot at the Shijingshan center is paired with a dedicated trainer who teleoperates it with VR headset and controllers while joint angles and grip force are recorded; methods also include exoskeleton teleoperation and first-person capture. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)The Standard (Hong Kong) | Why | The Shijingshan center's general manager said fewer than 5 percent of the 23,000 humanoids China shipped in the first half of 2026 could do practical factory work, largely because of a shortage of high-quality real-world training data. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)Global Times | Hours collected | At the Shijingshan center an eight-hour shift first yielded about 0.9 hours of usable data per data collector; after process upgrades the yield rose to five to six hours a day, per the center's figures cited by an analyst. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Facility | Leju Robotics says it has helped build nine humanoid robot training facilities across China, including the Shijingshan center. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Hiring drive | Leju's Shijingshan trainers need no advanced education: a background in mechanics or computers and physical stamina help, a new trainer works independently after one or two days of training, and a mature employee takes about a month. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Hours collected | The Shijingshan center can produce more than 8 million multimodal data records a year, and Leju says it produces nearly 25 million real-machine data records a year across the facilities it helped build. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)China Daily | Team size | The Shijingshan humanoid data-training center co-built by Leju Robotics and the Shijingshan district government had an instructor team of more than 200 people, still expanding, per Leju's training-center technical director. |
| Leju Robotics (Shijingshan Humanoid Robot Data Training Center, Beijing)Rest of World | Facility | The Shijingshan center, described as China's largest humanoid robot data-training center, covers more than 10,000 square metres with 16 scenarios including a car assembly line, a smart home and an elder-care facility. |
| LG ElectronicsThe Investor (Korea Herald) | Facility | LG's Yangjae data factory in Seoul, due for full operation by end 2026, has CLOiD robots collecting data in a mock home and in a replica of LG's Tennessee washing machine plant, plus spaces for LG CNS logistics and LG Innotek robotic hands. |
| LG ElectronicsThe Investor (Korea Herald) | Hours collected | LG Electronics expects to collect 100,000 hours of robot training data with its own CLOiD robots at its Yangjae data factory in Seoul by the end of 2026. |
| LG ElectronicsThe Investor (Korea Herald) | Why | LG plans to use the Yangjae data to advance its own robot foundation model. |
| Liuzhou humanoid robot training centre (Guangxi; UBTech hardware)New Straits Times (Reuters) | Facility | A government training centre in Liuzhou, Guangxi, holds more than 100 humanoid robots, supplied by UBTech under an 18 million US dollar regional tender, where a dozen headset-wearing trainers guide robots through tasks such as sorting crates and making coffee. |
| Liuzhou humanoid robot training centre (Guangxi; UBTech hardware)New Straits Times (Reuters) | Method | At the Liuzhou centre a novice trainer gets about one usable movement in 300 attempts and an experienced trainer about one in 50; trainers see through the robot's eyes via headsets and drive it with sensor-fitted handheld controllers. |
| Liuzhou humanoid robot training centre (Guangxi; UBTech hardware)New Straits Times (Reuters) | Why | The Liuzhou centre aims to sell robot-training data to factories, but three staff said the subsidy-dependent project had no clear path to profitability because of high operating costs and low data prices. |
| MotionalMotional (Greenhouse job board) | Hiring drive | Motional is hiring Autonomous Driving Test Operators for its growing AV Test Operations team (posting listed in Pittsburgh) to operate autonomous vehicles, collect and annotate data and give engineers feedback; the posted base salary is $52,000. |
| National and Local Co-Built Humanoid Robotics Innovation Center (Shanghai)Pudong New Area government (China Daily) | Facility | Shanghai's first heterogeneous humanoid robot training ground (formerly the Kylin Training Ground), run under the National and Local Co-Built Humanoid Robotics Innovation Center, opened in January 2025 on 4,600 square metres with more than 100 robots in 10 scenarios. |
| National and Local Co-Built Humanoid Robotics Innovation Center (Shanghai)Yicai Global | Why | The center's chief scientist called the Shanghai training base a factory that produces data, built to pool data from many robot makers because training on data from a single kind of robot creates an AI illusion. |
| NuroNuro (Greenhouse job board) | Pay | Nuro's Autonomous Vehicle Operator role in South San Francisco (12-month temporary assignment) pays $27 an hour on morning/afternoon shifts and $28.25 overnight; operators drive test vehicles in manual and autonomous modes up to eight hours a day and collect real-world driving data. |
| OpenAIBusiness Insider | Facility | Business Insider reported that OpenAI runs three shifts and a few dozen workstations collecting robot data around the clock in San Francisco. |
| OpenAIBusiness Insider | Hiring drive | OpenAI told employees it plans a second robotics data lab in Richmond, California, and a December job posting for a robotics operator with its contracting agency lists Richmond, Business Insider reported. |
| OpenAIBusiness Insider | Method | OpenAI's data collectors use 3D-printed GELLO controllers to teleoperate two Franka robot arms, and are rated on how many good hours of functional training data they generate, Business Insider reported. |
| OpenAIBusiness Insider | Team size | Business Insider reported that OpenAI's robotics lab in San Francisco employs around 100 data collectors and has more than quadrupled since it launched in February 2025. |
| Physical IntelligenceAshby (Physical Intelligence careers) | Facility | Physical Intelligence runs a Set Team that builds and maintains the physical rooms where its robots collect data, such as kitchens, living rooms and evaluation stations. |
| Physical IntelligenceAshby (Physical Intelligence careers) | Hiring drive | Physical Intelligence is hiring an operations recruiter for high-volume Robot Operator hiring, saying its operators, copilots, shift leads, technicians and props staff collect training data on-site in San Francisco and the teams are growing quickly. |
| Physical IntelligenceAshby (Physical Intelligence careers) | Method | Physical Intelligence Robot Operators either lead the robot with their own arms (the robot mirrors them) or use specialized controllers, and the job is metrics-based with data collection goals per shift. |
| Physical IntelligenceAshby (Physical Intelligence careers) | Pay | Physical Intelligence advertises an on-site Robot Operator role in San Francisco that teleoperates robot arms to collect training data, paying $25 an hour on 8-hour evening or overnight shifts, minimum 5 days a week. |
| Physical IntelligencearXiv (Physical Intelligence) | Hours collected | For pi0.5, Physical Intelligence used about 400 hours of data from its mobile manipulators doing household tasks in about 100 different homes. |
| Physical IntelligencearXiv (Physical Intelligence) | Hours collected | Physical Intelligence pre-trained its pi0 model on about 10,000 hours of dexterous manipulation data from its own robots (7 robot configurations, 68 tasks), on top of open datasets such as OXE, DROID and Bridge. |
| RealMan RoboticsElectropages | Facility | A Beijing Humanoid Robotics Data Training Centre where RealMan Robotics is the core equipment provider and operator opened on 3,000 square metres with 108 robots and ten real-world scenes, estimated to produce over one million data points a year. |
| Skild AISkild AI | Method | Skild AI said it is scaling robot teleoperation, UMI handheld grippers, egocentric video and simulation data collection in-house, because no single source wins on scalability, diversity and closeness to the robot. |
| Skild AISkild AI | Why | Skild AI says teleoperation alone cannot reach foundation-model scale, because it is tied to physical presence and happens in real time, so it trains on human video as well. |
| SkydioSkydio (Ashby job board) | Hiring drive | Skydio is hiring a fixed-term (12-month) Data Collection Flight Test Operator in San Mateo, California to fly its drones across varied US environments, about 90% of the time on the road, gathering the data that trains and validates its autonomy and AI. |
| SkydioSkydio (Ashby job board) | Pay | Skydio's Data Collection Flight Test Operator role pays $50 to $60 an hour. |
| Sunday RoboticsSunday Robotics | Method | Sunday Robotics trains its Memo robot on data from Memory Developers wearing its Skill Capture Glove at home, and says it has shipped thousands of gloves to contributors across the country. |
| Sunday RoboticsAshby (Sunday Robotics careers) | Pay | Sunday Robotics pays remote Memory Developers, who record household chores with its hardware, $30 per hour of approved data during a one-week screening and up to $60 per hour for complex tasks afterwards. |
| Sunday RoboticsAshby (Sunday Robotics careers) | Pay | Sunday Robotics also hires in-house Memory Developers at Redwood City at $35 an hour on a fixed 5-day, 8-hour evening shift (2 PM to 10 PM), who also pilot new tasks and hardware before they go to the remote collectors. |
| Sunday RoboticsSunday Robotics | Team size | Sunday Robotics said that ahead of its beta program it had grown Engineering 3x, Research 4x and Data Operations 5x. |
| Sunday RoboticsSunday Robotics | Why | Sunday Robotics said its ACT-1 model was trained without any teleoperation data, arguing that gathering data for a general-purpose robot by teleoperation would take decades. |
| TeslaAUTO BILD | Method | Tesla told staff at Gigafactory Berlin-Brandenburg (Gruenheide) that selected teams would wear cameras during their work so the footage can train Optimus, per Handelsblatt's report of a company email. |
| TeslaAUTO BILD | Pay | Auto Bild reported that Tesla's US Optimus Data Collection Operators are paid hourly wages between 25.25 and 34.50 dollars. |
| TeslaBusiness Insider | Facility | Tesla trains Optimus in a glass-walled lab at its engineering headquarters, where dozens of workers act out everyday motions such as lifting a cup or wiping a table. |
| TeslaBusiness Insider | Hours collected | Each Optimus data collector is expected to gather at least four hours of usable video footage per shift, five current and former workers said. |
| TeslaBusiness Insider | Method | Optimus data collectors work eight-hour shifts repeating each action hundreds of times, recorded by five cameras on a helmet and a heavy backpack (cameras Tesla made in-house). |
| TeslaBusiness Insider (via AOL) | Pay | Tesla's 'Vehicle Operator, Autopilot' listing in Queens, New York, for operators who drive engineering cars and collect audio and camera data, paid Test Operator I $25.25 to $27.60 an hour base and Test Operator II $28.75 to $30.60, with 10% more on afternoon and night shifts (up to $33.66). |
| TeslaBusiness Insider | Method | In late June 2025 Tesla told Optimus staff it would shift to a vision-only approach, recording videos of workers doing tasks instead of relying on motion-capture suits and teleoperation. |
| TeslaBusiness Insider | Why | Tesla told the Optimus team that moving away from motion-capture suits and teleoperation would let it scale data collection more quickly. |
| TeslaBusiness Insider | Team size | Tesla registered 224 in-house test drivers and 104 vehicles with the California DMV in December 2024 for its autonomous testing permit, up from 59 drivers and 14 vehicles, according to public records viewed by Business Insider. |
| TeslaBusiness Insider | Why | Tesla's 'Project Rodeo' test drivers are trained to let the self-driving software keep driving as long as possible; Tesla engineers told Business Insider the reason is that longer autonomous stretches give them more data to work with. |
| TeslaBusiness Insider | Method | Tesla's Optimus data collection operators walked test routes and performed designated tasks wearing a motion-capture suit and a VR headset for more than seven hours a day. |
| TeslaBusiness Insider | Pay | Tesla's Optimus 'Data Collection Operator' job listing offered full-time pay of $25.25 to $48 per hour. |
| TeslaBusiness Insider | Team size | More than 50 different people worked as Optimus data collection operators at Tesla over the year to August 2024, per a Business Insider review of LinkedIn data. |
| Toyota Research InstitutearXiv (Toyota Research Institute) | Facility | Of TRI's in-house data, 468 hours (362 tasks, 46,063 demonstrations) came from teleoperation across 9 robot hardware stations, plus 32 hours captured with 7 pairs of handheld UMI devices in the wild. |
| Toyota Research InstitutearXiv (Toyota Research Institute) | Hours collected | Toyota Research Institute trained its Large Behavior Models on about 1,700 hours of demonstrations, of which about 545 hours (532 tasks, 64,262 demonstrations) were collected in-house at TRI. |
| Toyota Research InstituteToyota USA Newsroom | Method | Toyota Research Institute said its robots learn new skills from haptic demonstrations by a human teacher using Diffusion Policy, and that it had taught over 60 skills with a target of 1,000 by the end of 2024. |
| UberTechCrunch | Hours collected | Uber said its planned 500-vehicle AV Labs fleet will be able to collect 2 million miles of high-fidelity driving data per month (a stated capacity, not miles already collected). |
| UberTechCrunch | Method | Uber said it plans to roll out 500 data-collection Hyundai Ioniq 5 vehicles globally in 2026 (50 on the road by summer), each fitted with 14 cameras, eight solid-state lidars and nine radars, to gather training data for its autonomous vehicle partners. |
| UberTechCrunch | Method | Uber launched an AV Labs division that sends its own sensor-equipped cars into cities to collect driving data for its autonomous vehicle partners (named examples: Waymo, Waabi, Lucid Motors); no partner contracts were signed at launch. |
| UberTechCrunch | Why | Uber's AV Labs head Danny Guo said partners want any data Uber can provide because Uber can collect far more driving data than the partners can gather with their own fleets. |
| WaabiWaabi (Lever job board) | Pay | Waabi's Vehicle Operator (CDL) day-shift role in Dallas, Texas, which manually drives Class 8 trucks on mapping and shadow missions and supervises the Waabi Driver, pays $31 to $43 an hour. |
| WaabiWaabi (Lever job board) | Why | Waabi says its self-driving system improves because its vehicle operators report precisely what they saw and why they took control, calling that feedback loop the core of the operator job. |
| WaymoCNBC (via NBC New York) | Method | Waymo confirmed it would test in Philadelphia through fall 2025 with a limited fleet and human safety drivers, first driving manually through downtown and freeways; CNBC reports the road trips collect mapping data and test the Waymo Driver in new environments. |
| WaymoWaymo | Method | On Waymo 'road trips', the company brings a small fleet of vehicles equipped with the Waymo Driver to a new city, where human specialists first drive them manually to give the system context and report local driving nuances to engineers. |
| WaymoWaymo | Why | Waymo says driving in dozens of cities has informed the design of its sensors, improved the Waymo Driver's performance in its existing cities and helped it bring the technology to new places. |
| WayveWayve (Ashby job board) | Hiring drive | Wayve is hiring Vehicle Safety Operators in Tokyo to monitor its AI driving system on public roads, intervene when required and support data collection that trains its AI. |
| WayveWayve (Ashby job board) | Why | Wayve describes on-road data collection as expensive and time-consuming, and says its synthetic data team exists to reduce that dependency by turning generative world models such as GAIA-3 into a source of training data. |
| ZiplineZipline (Greenhouse job board) | Facility | Zipline runs a flight test facility in Amarillo, Texas staffed by Flight Test Operator, Ground Test and Visual Observer teams, where aircraft move from engineering development into flight test; it is hiring a Site Lead to run it. |
| ZiplineZipline (Greenhouse job board) | Pay | Zipline's Flight Test Operator role at its Esparto, California test site, acting as Remote Pilot in Command for high-volume flight testing including nights and weekends, starts at $34 to $38 an hour. |
| ZiplineZipline (Greenhouse job board) | Why | Zipline says its high-volume flight testing exposes risks before they reach customers, strengthens the safety case for its aircraft and improves the reliability of every delivery. |
| ZooxAutomotive World | Method | Zoox said it would start testing in Dallas and Phoenix with retrofitted Toyota Highlander SUVs and safety drivers that map roads and collect operational data, taking its US testing footprint to ten cities. |
Base: All 115 facts.
Credit: compiled by RobotTraining.jobs Research, How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay, checked 18 Oct 2026.
Robotics labs · Fleet operators · Data vendors · Research teams
Buy the data behind these reports
- employer job boards read every day
- 103employer job boards read every day
- every role's history kept since
- 12 Sep 2026every role's history kept since
- apply clicks in the last 30 days
- 392apply clicks in the last 30 days
The raw records
Every robot training role we track: title, employer, place, kind of work, pay as stated, and the days it was first and last open.
A regular feed
New and closed roles every week or month, as file drops or through the data endpoints.
A custom cut
Any figure here by country, city, kind of work, vehicle, employer or date, or your own hiring benchmarked against the market.
Each report's own figures stay free to download and reuse under CC BY 4.0.
How it was measured
The companies' own posts, job postings and filings, and reporting by major outlets, each linked beside its fact with a verbatim quote kept in the dataset. Each row carries a link to its publication and gives the number exactly as published, with no estimates or conversions. A note under every figure says what it covers.
- In-house
- Data a company collects with people it employs or contracts directly, in its own facilities or vehicles, rather than buying it from a vendor.
- Stated or reported
- Stated: the company, a filing or its own posting says it. Reported: an outlet reports it from sources the company has not confirmed.
Limits of this data
- Built only from public sources. Anything never announced or reported is absent, and private-company figures are whatever the company or the press said, unaudited.
- Companies disclose unevenly, so the number of facts reflects what is public, not the size of an operation.
- Pay from job postings is the offer at the time it was read and can change.
Collection and classification of listings are explained on the methodology page. Freshness, completeness and consistency are tested daily in the data quality report.
Monthly editions
The live page shows the dataset as of its latest review, and new events join it once they are made public. Every finished month also gets its own edition, locked two days after month end, so a number quoted from it stays fixed.
- Live pageLatest review
Quoting this report
You may quote, chart or republish these figures under CC BY 4.0, provided you link back to this page. The live numbers shift every day. For a figure that stays put, quote a monthly edition, or note the date you accessed this page.
Citation
RobotTraining.jobs (2026). How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay (live data, 19 Sep 2023 to 18 Oct 2026). https://robottraining.jobs/research/robot-labs-in-house-data-collection (accessed 18 Oct 2026)
BibTeX
@misc{robottrainingjobs_robot_labs_in_house_data_collection_live,
author = {{RobotTraining.jobs}},
title = {How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and Pay (live data, 19 Sep 2023 to 18 Oct 2026)},
year = {2026},
howpublished = {\url{https://robottraining.jobs/research/robot-labs-in-house-data-collection}},
note = {Accessed 18 Oct 2026},
}Files: every figure in one CSV (a row per value) or as JSON. Each figure can also be downloaded separately. Other open data: datasets and API.
Keep reading
- Funding and Valuations of the Companies That Train RobotsHow much money has gone into the companies that hire people to train and operate robots, and at what valuations?Open report
- Revenue and Scale of the Companies That Run RobotsHow much revenue do the listed companies that operate delivery robots and autonomous vehicles book, and how large are their operations?Open report
- Robot Operator Legal Tracker: Rules, Inquiries, Rulings and LayoffsWhich laws, permits, inquiries, rulings and layoffs concern the people who operate and supervise robots and autonomous vehicles, and where do they stand?Open report
- Who Is Buying Robot Training Data: Deals, Paid Programs and Dataset ReleasesWhich companies pay for, license, commission or release training data for robots, embodied AI, autonomous vehicles and drones, from whom, and what kind of data?Open report
- Robot Training Data Vendors: Who Sells It, to Whom, and What They PayWhich companies sell robot training data or the human work behind it (teleoperation, demonstrations, egocentric capture, annotation, AV and drone operation), who are their customers, how big are they and what do they pay workers?Open report
- Pay in Robot Data and Operator Work: What Reporting Has FoundWhat have reporters and researchers documented about the pay and conditions of robot data collectors and teleoperators, and what do open roles here advertise?Open report
- Robot Training Pay Report 2026How much do robot training jobs pay, broken down by kind of work and employer?Open report