Skip to content

Research · October 2026 edition

How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video

50 robot foundation models and the data they were trained on: hours and episodes, how much people made by teleoperating robots or filming their own hands, and how many robot types; 79 facts, each linked to its source.

Read the month's story: Robot Training Jobs in October 2026: The Month in Data

Not final yet. The month has ended, yet stragglers in the data may still land up to 3 Nov 2026; after that the edition is locked permanently.

Covers
20 Jun 2023 to 28 Oct 2026
Dated by when each event occurred
Calculated
1 Nov 2026, 06:05 UTC
Settles when locked
Where it comes from
Developers' papers, model cards and announcements, and reporting by major outlets where a developer has not published, each linked with a verbatim quote kept in the dataset.
Method
Revision 1. Terms defined
CC BY 4.0, citing this report

Each figure here comes as CSV and JSON files, emailed to your work address.

Need robot training data collected? We run the project: teleoperation sessions, first-person video, filmed household tasks or labeled robot video, staffed from the people who apply for this work through us, checked against your standard, starting with a paid pilot.

How we run projects

What the data shows

  1. 1

    GR00T N1 is the most documented of the 50 robot models (4 sourced facts).

    Figure 1
  2. 2

    GEN-1 (Generalist AI) states the most hours of training data on record: 500,000 hours.

    Figure 2
  3. 3

    47 statements from 34 models say how much of their data came from teleoperation, human video or simulation.

    Figure 3
  4. 4

    The most recent fact on record (28 Oct 2026): AgiBot debuted GO-3 Preview, its next-generation embodied foundation model, at AGIBOT Innovation Day 2026 in Shanghai (the press release is stamped 28 October 2026).

    Figure 4

Figure 1

Robot foundation models and the data behind them

How much human-made data was each robot foundation model trained on, and where did it come from?

ModelTraining data as statedSourced facts
GR00T N1arXiv (NVIDIA)NVIDIAGR00T N1 is trained on a mixture of real robot trajectories, human videos and synthetic data, arranged as a pyramid with web data and human video at the base and real robot data at the top.NVIDIA trained GR00T N1 on 88 hours of in-house teleoperation data on the Fourier GR-1 humanoid, which it expanded to 827 hours of generated video with video models.n/a
Large Behavior Model (LBM)arXiv (Toyota Research Institute)Toyota Research InstituteTRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data.The real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.n/a
G0arXiv (Galaxea)GalaxeaThe Galaxea Open-World Dataset used to train G0 holds 500 hours of robot data recorded in real homes and workplaces, over 150 tasks in 50 scenes.Galaxea recorded all of its dataset's demonstrations on the Galaxea R1 Lite robot with an isomorphic teleoperation scheme that maps the operator's movements directly to the robot.n/a
GEN-1Generalist AIGeneralist AIGeneralist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.Generalist says the base models of GEN-0 and GEN-1 use no robot data at all: they are pre-trained on data from low-cost wearable devices worn by humans doing millions of activities.n/a
GO-1 (Genie Operator-1)arXiv (AgiBot)AgiBotThe AgiBot World dataset behind GO-1 holds 1,001,552 trajectories totalling 2,976.4 hours across 217 tasks, collected by more than 100 identical robots.AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.n/a
GR-3arXiv (ByteDance Seed)ByteDance SeedByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.n/a
π0arXiv (Physical Intelligence)Physical IntelligencePhysical Intelligence pre-trained π0 on about 10,000 hours of its own dexterous robot manipulation data, alongside open datasets such as Open X-Embodiment.n/aPhysical Intelligence released the code and weights of π0 as open source on February 4, 2025, saying 1 to 20 hours of data had been enough to fine-tune it to a variety of tasks.
Skild BrainSkild AISkild AISkild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.Skild AI said it fine-tunes its model to perform new skills by watching human videos, plus a small amount of robot data that the post puts at under one hour.n/a
1X World Model (1XWM)1X Technologies1X Technologies1X fine-tuned 1XWM on 70 hours of robot data to adapt it to NEO's appearance and kinematics, saying the model needs no large-scale robot pre-training or teleoperated demonstrations of the tasks.1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.n/a
ACT-1Sunday RoboticsSunday RoboticsSunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.n/a
Dyna-2Dyna RoboticsDyna Roboticsn/aDyna Robotics pre-trained Dyna-2 on more than one million hours of egocentric human video, mostly head-mounted recordings of everyday manipulation, with no robot data in pre-training.n/a
GEN-0Generalist AIGeneralist AIGeneralist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week.n/an/a
GR00T N1.5NVIDIA Research (GEAR)NVIDIAGR00T N1.5's pre-training mixture combined NVIDIA's internal GR-1 humanoid data, Open X-Embodiment, simulated GR-1 data, DreamGen synthetic trajectories and AgiBot-Beta data.n/an/a
Helix 2.5FigureFigure AIn/aFigure pretrained Helix 2.5 on Index, its global-scale dataset of human behavior, then ran three behaviors in 30 Bay Area homes with zero data collected in any of them.n/a
OptimusBusiness Insider (via AOL)Teslan/aBusiness Insider reported, citing insiders, that Tesla would train Optimus mainly on videos of workers doing tasks rather than on motion capture suits and teleoperation, to scale data collection faster.n/a
π0.5arXiv (Physical Intelligence)Physical Intelligenceπ0.5 uses about 400 hours of data from mobile manipulators doing household tasks in about 100 different homes, the slice of its training data closest to its test tasks.n/an/a
RFM-1MIT Technology ReviewCovariantMIT Technology Review reported that RFM-1 was trained on years of data from Covariant's warehouse item-picking robots plus words and videos from the internet.n/an/a
RT-X (Open X-Embodiment)Google DeepMindGoogle DeepMindGoogle DeepMind said the Open X-Embodiment dataset used to train RT-1-X and RT-2-X spans more than 1 million episodes, 500 skills and 150,000 tasks.n/an/a
SmolVLAarXiv (Hugging Face)Hugging FaceHugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets.n/an/a
ACT-2Sunday RoboticsSunday RoboticsSunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.n/an/a
Atlas Large Behavior ModelBoston DynamicsBoston Dynamics and Toyota Research Instituten/aBoston Dynamics collects the training data for Atlas's large behavior model with a VR-based teleoperation system, on the real robot and in simulation.n/a
Diffusion Policy behavior modelToyota USA NewsroomToyota Research Instituten/aTRI said in 2023 that its robot behavior model learns from haptic demonstrations by a human teacher plus a language description of the goal, and that it had taught robots over 60 skills this way.n/a
Gemini RoboticsarXiv (Google DeepMind)Google DeepMindn/aGoogle DeepMind collected a teleoperated robot action dataset on a fleet of ALOHA 2 robots over 12 months, amounting to thousands of hours of expert demonstrations, to train Gemini Robotics.n/a
Gemini Robotics 1.5arXiv (Google DeepMind)Google DeepMindn/an/an/a
Gemini Robotics 2ForkLogGoogle DeepMindn/an/aGoogle DeepMind introduced Gemini Robotics 2, a vision-language-action model for full-body robot control, in late July 2026, with an embodied reasoning model (ER 2) and an on-device model alongside it.
Gemini Robotics On-DeviceGoogle DeepMindGoogle DeepMindn/an/an/a
Gemini Robotics On-Device 2ForkLogGoogle DeepMindForkLog reported that Gemini Robotics On-Device 2 adapts to new robots with fewer than 200 examples in just a few hours of training; a second outlet reports the same figure from DeepMind.n/an/a
GEN-1.5Generalist AIGeneralist AIn/an/aGeneralist released GEN-1.5 on 19 August 2026, a robot foundation model that can be prompted in context with one demonstration recorded as human data with handheld grippers or as robot rollouts.
GO-3 PreviewAgiBotAgiBotn/an/aAgiBot debuted GO-3 Preview, its next-generation embodied foundation model, at AGIBOT Innovation Day 2026 in Shanghai (the press release is stamped 28 October 2026).
GR-2arXiv (ByteDance Research)ByteDance Seedn/aByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.n/a
GR00T N1.6NVIDIA Research (GEAR)NVIDIAn/aOn top of the N1.5 mixture, GR00T N1.6 adds several thousand hours of teleoperated data from bimanual YAM arms, AgiBot Genie-1, a simulated Galaxea R1 Pro and whole-body Unitree G1 locomanipulation.n/a
GraspVLAarXiv (Galbot, Peking University)GalbotGalbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.n/an/a
HelixFigure AIFigure AIn/aFigure trained Helix on about 500 hours of teleoperated behaviors collected across multiple robots and operators, which it said is a small fraction of the size of earlier VLA datasets.n/a
Helix 02Figure AIFigure AIHelix 02's whole-body controller (System 0) was trained on over 1,000 hours of retargeted human motion data, with reinforcement learning across more than 200,000 simulated environments.n/an/a
Helix (Project Go-Big)Figure AIFigure AIn/aFigure said Helix learned to navigate cluttered spaces from spoken commands after training only on egocentric human video, collected under its Project Go-Big program including in Brookfield properties.n/a
λ0 (HumanVerse)arXiv (Lambda-humanoid Team)Lambda-humanoid Teamn/aThe Lambda-humanoid Team introduced HumanVerse, a 500-hour dataset of open-world human loco-manipulation pairing egocentric video with synchronized body and hand motion, to train its λ0 humanoid model.n/a
OctoarXivUC Berkeley, Stanford, CMU and Google DeepMind (Octo Model Team)Octo was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection.n/an/a
OpenVLAarXivStanford, UC Berkeley, Toyota Research Institute, Google DeepMind and othersOpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset.n/an/a
π*0.6Physical IntelligencePhysical Intelligencen/aPhysical Intelligence says robot foundation models today mainly learn from demonstrations people collect by hand, such as teleoperation; for π*0.6 it adds expert teleoperator corrections and the robot's own autonomous experience.n/a
π0.5 (human-to-robot transfer)Physical IntelligencePhysical Intelligencen/aPhysical Intelligence found that egocentric video of humans, recorded with wearable cameras, roughly doubled performance on tasks where robot data is limited once its models were trained on enough robot data.n/a
π0.7Physical IntelligencePhysical Intelligenceπ0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.n/an/a
pi05_hsr_base (AIRoA)AIRoAAI Robot Association (AIRoA)n/an/aOn 31 July 2026 AIRoA publicly released about 5,000 hours of cleansed robot motion data covering 68 task types, plus a continually pretrained model based on the open-source pi0.5 and its training code.
RDT-1BRDT project (Tsinghua University)Tsinghua UniversityRDT-1B was pre-trained on 46 multi-robot datasets with over 1 million episodes, then fine-tuned on over 6,000 episodes collected on an ALOHA dual-arm robot.n/an/a
Redwood1X Technologies1X Technologiesn/a1X trained its Redwood model on a large dataset of teleoperated and autonomous episodes from its EVE and NEO robots, gathered in 1X offices and employee homes.n/a
RoboCatGoogle DeepMindGoogle DeepMindn/aTo teach RoboCat a new task or robot, DeepMind collected 100 to 1,000 demonstrations with a human-controlled robot arm, then let RoboCat generate more training data itself.n/a
RT-2Google DeepMindGoogle DeepMindRT-2 was co-fine-tuned on web data and on RT-1 robot demonstration data collected with 13 robots over 17 months in an office kitchen.n/an/a
Skild S1Skild AISkild AIFor S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours.n/an/a
V-JEPA 2arXiv (Meta FAIR)Metan/aMeta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.n/a
V-JEPA 2-ACarXiv (Meta FAIR)MetaMeta post-trained the action-conditioned V-JEPA 2-AC world model on less than 62 hours of unlabeled robot video from the DROID dataset.n/an/a
Xiaomi-Robotics-1arXiv (Xiaomi Robotics Team)XiaomiXiaomi's robotics team pre-trained its Xiaomi-Robotics-1 vision-language-action model on over 100,000 hours of real-world manipulation trajectories collected with UMI devices.n/an/a

Base: 50 robot models, 79 facts, each linked to the page that states it (reviewed 28 Oct 2026).

A model appears once its developer or a major outlet has said something about its training data.

Credit: compiled by RobotTraining.jobs Research, How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition.

Figure 2

Models by stated hours of training data

Which robot models were trained on the most hours of data?

ModelHours
GEN-1Generalist AIGeneralist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.2 Apr 2026
GEN-0Generalist AIGeneralist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week.4 Nov 2025
Xiaomi-Robotics-1arXiv (Xiaomi Robotics Team)Xiaomi's robotics team pre-trained its Xiaomi-Robotics-1 vision-language-action model on over 100,000 hours of real-world manipulation trajectories collected with UMI devices.16 Jul 2026
π0arXiv (Physical Intelligence)Physical Intelligence pre-trained π0 on about 10,000 hours of its own dexterous robot manipulation data, alongside open datasets such as Open X-Embodiment.31 Oct 2024
Large Behavior Model (LBM)arXiv (Toyota Research Institute)TRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data.7 Jul 2025
Helix 02Figure AIHelix 02's whole-body controller (System 0) was trained on over 1,000 hours of retargeted human motion data, with reinforcement learning across more than 200,000 simulated environments.27 Jan 2026
G0arXiv (Galaxea)The Galaxea Open-World Dataset used to train G0 holds 500 hours of robot data recorded in real homes and workplaces, over 150 tasks in 50 scenes.30 Aug 2025
π0.5arXiv (Physical Intelligence)π0.5 uses about 400 hours of data from mobile manipulators doing household tasks in about 100 different homes, the slice of its training data closest to its test tasks.22 Apr 2025
1X World Model (1XWM)1X Technologies1X fine-tuned 1XWM on 70 hours of robot data to adapt it to NEO's appearance and kinematics, saying the model needs no large-scale robot pre-training or teleoperated demonstrations of the tasks.12 Jan 2026
V-JEPA 2-ACarXiv (Meta FAIR)Meta post-trained the action-conditioned V-JEPA 2-AC world model on less than 62 hours of unlabeled robot video from the DROID dataset.11 Jun 2025

Base: 10 robot models with a stated figure; each row is the largest one on record for it.

Hours as each developer counts them (robot data, human video or both); they are not comparable one to one.

Credit: compiled by RobotTraining.jobs Research, How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition.

Figure 3

How much of it people made by teleoperating robots or filming themselves

What share of robot model training data comes from teleoperation and from human video?

ModelAs stated
λ0 (HumanVerse)arXiv (Lambda-humanoid Team)The Lambda-humanoid Team introduced HumanVerse, a 500-hour dataset of open-world human loco-manipulation pairing egocentric video with synchronized body and hand motion, to train its λ0 humanoid model.13 Oct 2026
Helix 2.5FigureFigure pretrained Helix 2.5 on Index, its global-scale dataset of human behavior, then ran three behaviors in 30 Bay Area homes with zero data collected in any of them.17 Sep 2026
Helix 2.5FigureFigure said its Index human-behavior dataset is now generating roughly 35 minutes of new human experience every second.17 Sep 2026
Dyna-2Dyna RoboticsDyna Robotics pre-trained Dyna-2 on more than one million hours of egocentric human video, mostly head-mounted recordings of everyday manipulation, with no robot data in pre-training.1 Aug 2026
Dyna-2Dyna RoboticsDyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands.1 Aug 2026
ACT-2Sunday RoboticsSunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.17 Jul 2026
π0.7Physical Intelligenceπ0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.16 Apr 2026
GEN-1Generalist AIGeneralist says the base models of GEN-0 and GEN-1 use no robot data at all: they are pre-trained on data from low-cost wearable devices worn by humans doing millions of activities.2 Apr 2026
1X World Model (1XWM)1X Technologies1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.12 Jan 2026
Skild BrainSkild AISkild AI said it fine-tunes its model to perform new skills by watching human videos, plus a small amount of robot data that the post puts at under one hour.12 Jan 2026
π0.5 (human-to-robot transfer)Physical IntelligencePhysical Intelligence found that egocentric video of humans, recorded with wearable cameras, roughly doubled performance on tasks where robot data is limited once its models were trained on enough robot data.16 Dec 2025
GR00T N1.6NVIDIA Research (GEAR)On top of the N1.5 mixture, GR00T N1.6 adds several thousand hours of teleoperated data from bimanual YAM arms, AgiBot Genie-1, a simulated Galaxea R1 Pro and whole-body Unitree G1 locomanipulation.15 Dec 2025
ACT-1Sunday RoboticsSunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.19 Nov 2025
ACT-1Sunday RoboticsSunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.19 Nov 2025
π*0.6Physical IntelligencePhysical Intelligence says robot foundation models today mainly learn from demonstrations people collect by hand, such as teleoperation; for π*0.6 it adds expert teleoperator corrections and the robot's own autonomous experience.17 Nov 2025
GEN-0Generalist AIGeneralist said GEN-0's data was collected in thousands of homes, warehouses and workplaces by thousands of data collection devices and robots, together with several data foundry partners.4 Nov 2025
Helix (Project Go-Big)Figure AIFigure said Helix learned to navigate cluttered spaces from spoken commands after training only on egocentric human video, collected under its Project Go-Big program including in Brookfield properties.18 Sep 2025
G0arXiv (Galaxea)Galaxea recorded all of its dataset's demonstrations on the Galaxea R1 Lite robot with an isomorphic teleoperation scheme that maps the operator's movements directly to the robot.30 Aug 2025
G0arXiv (Galaxea)G0's first pre-training stage uses about 1,000 hours of Open X-Embodiment trajectories, 500 hours from the Galaxea Open-World Dataset and 200 hours of in-house data.30 Aug 2025
OptimusBusiness Insider (via AOL)Business Insider reported, citing insiders, that Tesla would train Optimus mainly on videos of workers doing tasks rather than on motion capture suits and teleoperation, to scale data collection faster.25 Aug 2025
OptimusBusiness Insider (via AOL)Before the shift to video, Tesla recorded Optimus training data from human operators wearing motion capture suits and VR headsets, who also teleoperated the robot.25 Aug 2025
Atlas Large Behavior ModelBoston DynamicsBoston Dynamics collects the training data for Atlas's large behavior model with a VR-based teleoperation system, on the real robot and in simulation.20 Aug 2025
Skild BrainSkild AISkild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.29 Jul 2025
GR-3arXiv (ByteDance Seed)ByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.21 Jul 2025
GR-3arXiv (ByteDance Seed)For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.21 Jul 2025
GR-3arXiv (ByteDance Seed)ByteDance said human trajectories recorded with VR devices come in at about 450 per hour, against about 250 per hour for teleoperated robot trajectories.21 Jul 2025
Large Behavior Model (LBM)arXiv (Toyota Research Institute)The real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.7 Jul 2025
Large Behavior Model (LBM)arXiv (Toyota Research Institute)TRI's LBM data also includes 32 hours (10,851 demonstrations) recorded by people with handheld Universal Manipulation Interface devices in the wild, and 45 hours from simulation.7 Jul 2025
GR00T N1.5NVIDIA Research (GEAR)GR00T N1.5's pre-training mixture combined NVIDIA's internal GR-1 humanoid data, Open X-Embodiment, simulated GR-1 data, DreamGen synthetic trajectories and AgiBot-Beta data.11 Jun 2025
V-JEPA 2arXiv (Meta FAIR)Meta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.11 Jun 2025
Redwood1X Technologies1X trained its Redwood model on a large dataset of teleoperated and autonomous episodes from its EVE and NEO robots, gathered in 1X offices and employee homes.10 Jun 2025
SmolVLAHugging FaceSmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all.3 Jun 2025
GR00T N1.5NVIDIA NewsroomNVIDIA said it used its GR00T-Dreams blueprint to generate the synthetic training data for GR00T N1.5 in 36 hours, against nearly three months of manual human data collection.18 May 2025
GraspVLAarXiv (Galbot, Peking University)Galbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.6 May 2025
π0.5arXiv (Physical Intelligence)97.6% of the training examples in π0.5's first training phase do not come from mobile manipulators doing household tasks but from other robots and web data.22 Apr 2025
GR00T N1arXiv (NVIDIA)NVIDIA trained GR00T N1 on 88 hours of in-house teleoperation data on the Fourier GR-1 humanoid, which it expanded to 827 hours of generated video with video models.18 Mar 2025
GR00T N1arXiv (NVIDIA)GR00T N1 is trained on a mixture of real robot trajectories, human videos and synthetic data, arranged as a pyramid with web data and human video at the base and real robot data at the top.18 Mar 2025
GR00T N1arXiv (NVIDIA)The base of GR00T N1's training data is egocentric human video from open datasets (such as Ego4D, Ego-Exo4D, EPIC-KITCHENS and HOI4D) together with web data.18 Mar 2025
Gemini RoboticsarXiv (Google DeepMind)Google DeepMind collected a teleoperated robot action dataset on a fleet of ALOHA 2 robots over 12 months, amounting to thousands of hours of expert demonstrations, to train Gemini Robotics.12 Mar 2025
GO-1 (Genie Operator-1)arXiv (AgiBot)AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.10 Mar 2025
GO-1 (Genie Operator-1)Newsfile Corp (AgiBot release)AgiBot launched GO-1 on March 10, 2025; its latent planner learns from cross-embodiment and human operation data, and its action expert from over a million real robot demonstrations.10 Mar 2025
HelixFigure AIFigure trained Helix on about 500 hours of teleoperated behaviors collected across multiple robots and operators, which it said is a small fraction of the size of earlier VLA datasets.20 Feb 2025
GR-2arXiv (ByteDance Research)ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.8 Oct 2024
RFM-1MIT Technology ReviewMIT Technology Review reported that RFM-1 was trained on years of data from Covariant's warehouse item-picking robots plus words and videos from the internet.11 Mar 2024
Diffusion Policy behavior modelToyota USA NewsroomTRI said in 2023 that its robot behavior model learns from haptic demonstrations by a human teacher plus a language description of the goal, and that it had taught robots over 60 skills this way.19 Sep 2023
RT-2Google DeepMindRT-2 was co-fine-tuned on web data and on RT-1 robot demonstration data collected with 13 robots over 17 months in an office kitchen.28 Jul 2023
RoboCatGoogle DeepMindTo teach RoboCat a new task or robot, DeepMind collected 100 to 1,000 demonstrations with a human-controlled robot arm, then let RoboCat generate more training data itself.20 Jun 2023

Base: 47 facts from 34 robot models.

Credit: compiled by RobotTraining.jobs Research, How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition.

Figure 4

Every sourced fact

What exactly has been published about each model?

ModelKindFact
1X World Model (1XWM)1X TechnologiesHuman video1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.12 Jan 2026Stated by the company or an official page
1X World Model (1XWM)1X TechnologiesHours of training data1X fine-tuned 1XWM on 70 hours of robot data to adapt it to NEO's appearance and kinematics, saying the model needs no large-scale robot pre-training or teleoperated demonstrations of the tasks.12 Jan 2026Stated by the company or an official page
ACT-1Sunday RoboticsTeleoperation dataSunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.19 Nov 2025Stated by the company or an official page
ACT-1Sunday RoboticsData mixSunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.19 Nov 2025Stated by the company or an official page
ACT-2Sunday RoboticsData mixSunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.17 Jul 2026Stated by the company or an official page
Atlas Large Behavior ModelBoston DynamicsTeleoperation dataBoston Dynamics collects the training data for Atlas's large behavior model with a VR-based teleoperation system, on the real robot and in simulation.20 Aug 2025Stated by the company or an official page
Diffusion Policy behavior modelToyota USA NewsroomTeleoperation dataTRI said in 2023 that its robot behavior model learns from haptic demonstrations by a human teacher plus a language description of the goal, and that it had taught robots over 60 skills this way.19 Sep 2023Stated by the company or an official page
Dyna-2Dyna RoboticsHuman videoDyna Robotics pre-trained Dyna-2 on more than one million hours of egocentric human video, mostly head-mounted recordings of everyday manipulation, with no robot data in pre-training.1 Aug 2026Stated by the company or an official page
Dyna-2Dyna RoboticsTeleoperation dataDyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands.1 Aug 2026Stated by the company or an official page
G0arXiv (Galaxea)Hours of training dataThe Galaxea Open-World Dataset used to train G0 holds 500 hours of robot data recorded in real homes and workplaces, over 150 tasks in 50 scenes.30 Aug 2025Stated by the company or an official page
G0arXiv (Galaxea)Teleoperation dataGalaxea recorded all of its dataset's demonstrations on the Galaxea R1 Lite robot with an isomorphic teleoperation scheme that maps the operator's movements directly to the robot.30 Aug 2025Stated by the company or an official page
G0arXiv (Galaxea)Data mixG0's first pre-training stage uses about 1,000 hours of Open X-Embodiment trajectories, 500 hours from the Galaxea Open-World Dataset and 200 hours of in-house data.30 Aug 2025Stated by the company or an official page
Gemini RoboticsarXiv (Google DeepMind)Teleoperation dataGoogle DeepMind collected a teleoperated robot action dataset on a fleet of ALOHA 2 robots over 12 months, amounting to thousands of hours of expert demonstrations, to train Gemini Robotics.12 Mar 2025Stated by the company or an official page
Gemini Robotics 1.5arXiv (Google DeepMind)Robot typesGemini Robotics 1.5 is pre-trained on robot data from three platforms (ALOHA, bi-arm Franka and the Apollo humanoid) plus public text, image and video data, and controls all three without robot-specific post-training.25 Sep 2025Stated by the company or an official page
Gemini Robotics 2ForkLogReleaseGoogle DeepMind introduced Gemini Robotics 2, a vision-language-action model for full-body robot control, in late July 2026, with an embodied reasoning model (ER 2) and an on-device model alongside it.31 Jul 2026Reported by an outlet
Gemini Robotics On-DeviceGoogle DeepMindRobot typesGemini Robotics On-Device was trained only on ALOHA robots and then adapted to a bi-arm Franka FR3 and Apptronik's Apollo humanoid; it adapts to new tasks with as few as 50 to 100 demonstrations.24 Jun 2025Stated by the company or an official page
Gemini Robotics On-Device 2ForkLogEpisodesForkLog reported that Gemini Robotics On-Device 2 adapts to new robots with fewer than 200 examples in just a few hours of training; a second outlet reports the same figure from DeepMind.31 Jul 2026Reported by an outlet
GEN-0Generalist AIHours of training dataGeneralist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week.4 Nov 2025Stated by the company or an official page
GEN-0Generalist AIData mixGeneralist said GEN-0's data was collected in thousands of homes, warehouses and workplaces by thousands of data collection devices and robots, together with several data foundry partners.4 Nov 2025Stated by the company or an official page
GEN-1Generalist AIHours of training dataGeneralist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.2 Apr 2026Stated by the company or an official page
GEN-1Generalist AIHuman videoGeneralist says the base models of GEN-0 and GEN-1 use no robot data at all: they are pre-trained on data from low-cost wearable devices worn by humans doing millions of activities.2 Apr 2026Stated by the company or an official page
GEN-1Generalist AIHours of training dataEach GEN-1 task result Generalist showed was built with about one hour of robot data.2 Apr 2026Stated by the company or an official page
GEN-1.5Generalist AIReleaseGeneralist released GEN-1.5 on 19 August 2026, a robot foundation model that can be prompted in context with one demonstration recorded as human data with handheld grippers or as robot rollouts.19 Aug 2026Stated by the company or an official page
GO-1 (Genie Operator-1)arXiv (AgiBot)EpisodesThe AgiBot World dataset behind GO-1 holds 1,001,552 trajectories totalling 2,976.4 hours across 217 tasks, collected by more than 100 identical robots.10 Mar 2025Stated by the company or an official page
GO-1 (Genie Operator-1)arXiv (AgiBot)Teleoperation dataAgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.10 Mar 2025Stated by the company or an official page
GO-1 (Genie Operator-1)Newsfile Corp (AgiBot release)Data mixAgiBot launched GO-1 on March 10, 2025; its latent planner learns from cross-embodiment and human operation data, and its action expert from over a million real robot demonstrations.10 Mar 2025Stated by the company or an official page
GO-3 PreviewAgiBotReleaseAgiBot debuted GO-3 Preview, its next-generation embodied foundation model, at AGIBOT Innovation Day 2026 in Shanghai (the press release is stamped 28 October 2026).28 Oct 2026Stated by the company or an official page
GR-2arXiv (ByteDance Research)Human videoByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.8 Oct 2024Stated by the company or an official page
GR-3arXiv (ByteDance Seed)Data mixByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.21 Jul 2025Stated by the company or an official page
GR-3arXiv (ByteDance Seed)Teleoperation dataFor one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.21 Jul 2025Stated by the company or an official page
GR-3arXiv (ByteDance Seed)Data mixByteDance said human trajectories recorded with VR devices come in at about 450 per hour, against about 250 per hour for teleoperated robot trajectories.21 Jul 2025Stated by the company or an official page
GR00T N1arXiv (NVIDIA)Teleoperation dataNVIDIA trained GR00T N1 on 88 hours of in-house teleoperation data on the Fourier GR-1 humanoid, which it expanded to 827 hours of generated video with video models.18 Mar 2025Stated by the company or an official page
GR00T N1arXiv (NVIDIA)Data mixGR00T N1 is trained on a mixture of real robot trajectories, human videos and synthetic data, arranged as a pyramid with web data and human video at the base and real robot data at the top.18 Mar 2025Stated by the company or an official page
GR00T N1arXiv (NVIDIA)EpisodesNVIDIA generated 780,000 simulated trajectories for GR00T N1's pre- and post-training, which it said equal 6,500 hours (nine continuous months) of human demonstration data.18 Mar 2025Stated by the company or an official page
GR00T N1arXiv (NVIDIA)Human videoThe base of GR00T N1's training data is egocentric human video from open datasets (such as Ego4D, Ego-Exo4D, EPIC-KITCHENS and HOI4D) together with web data.18 Mar 2025Stated by the company or an official page
GR00T N1.5NVIDIA Research (GEAR)Data mixGR00T N1.5's pre-training mixture combined NVIDIA's internal GR-1 humanoid data, Open X-Embodiment, simulated GR-1 data, DreamGen synthetic trajectories and AgiBot-Beta data.11 Jun 2025Stated by the company or an official page
GR00T N1.5NVIDIA NewsroomData mixNVIDIA said it used its GR00T-Dreams blueprint to generate the synthetic training data for GR00T N1.5 in 36 hours, against nearly three months of manual human data collection.18 May 2025Stated by the company or an official page
GR00T N1.6NVIDIA Research (GEAR)Teleoperation dataOn top of the N1.5 mixture, GR00T N1.6 adds several thousand hours of teleoperated data from bimanual YAM arms, AgiBot Genie-1, a simulated Galaxea R1 Pro and whole-body Unitree G1 locomanipulation.15 Dec 2025Stated by the company or an official page
GraspVLAarXiv (Galbot, Peking University)Data mixGalbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.6 May 2025Stated by the company or an official page
HelixFigure AITeleoperation dataFigure trained Helix on about 500 hours of teleoperated behaviors collected across multiple robots and operators, which it said is a small fraction of the size of earlier VLA datasets.20 Feb 2025Stated by the company or an official page
Helix (Project Go-Big)Figure AIHuman videoFigure said Helix learned to navigate cluttered spaces from spoken commands after training only on egocentric human video, collected under its Project Go-Big program including in Brookfield properties.18 Sep 2025Stated by the company or an official page
Helix 02Figure AIHours of training dataHelix 02's whole-body controller (System 0) was trained on over 1,000 hours of retargeted human motion data, with reinforcement learning across more than 200,000 simulated environments.27 Jan 2026Stated by the company or an official page
Helix 2.5FigureHuman videoFigure pretrained Helix 2.5 on Index, its global-scale dataset of human behavior, then ran three behaviors in 30 Bay Area homes with zero data collected in any of them.17 Sep 2026Stated by the company or an official page
Helix 2.5FigureHuman videoFigure said its Index human-behavior dataset is now generating roughly 35 minutes of new human experience every second.17 Sep 2026Stated by the company or an official page
Large Behavior Model (LBM)arXiv (Toyota Research Institute)Hours of training dataTRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data.7 Jul 2025Stated by the company or an official page
Large Behavior Model (LBM)arXiv (Toyota Research Institute)EpisodesTRI's own share of the LBM pre-training data (TRI-Ramen) is 545 hours over 532 tasks, totalling 64,262 demonstrations; the rest is about 1,150 hours of curated Open X-Embodiment data.7 Jul 2025Stated by the company or an official page
Large Behavior Model (LBM)arXiv (Toyota Research Institute)Teleoperation dataThe real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.7 Jul 2025Stated by the company or an official page
Large Behavior Model (LBM)arXiv (Toyota Research Institute)Data mixTRI's LBM data also includes 32 hours (10,851 demonstrations) recorded by people with handheld Universal Manipulation Interface devices in the wild, and 45 hours from simulation.7 Jul 2025Stated by the company or an official page
OctoarXivEpisodesOcto was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection.20 May 2024Stated by the company or an official page
OpenVLAarXivEpisodesOpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset.13 Jun 2024Stated by the company or an official page
OptimusBusiness Insider (via AOL)Human videoBusiness Insider reported, citing insiders, that Tesla would train Optimus mainly on videos of workers doing tasks rather than on motion capture suits and teleoperation, to scale data collection faster.25 Aug 2025Reported by an outlet
OptimusBusiness Insider (via AOL)Teleoperation dataBefore the shift to video, Tesla recorded Optimus training data from human operators wearing motion capture suits and VR headsets, who also teleoperated the robot.25 Aug 2025Reported by an outlet
pi05_hsr_base (AIRoA)AIRoAReleaseOn 31 July 2026 AIRoA publicly released about 5,000 hours of cleansed robot motion data covering 68 task types, plus a continually pretrained model based on the open-source pi0.5 and its training code.31 Jul 2026Stated by the company or an official page
RDT-1BRDT project (Tsinghua University)EpisodesRDT-1B was pre-trained on 46 multi-robot datasets with over 1 million episodes, then fine-tuned on over 6,000 episodes collected on an ALOHA dual-arm robot.10 Oct 2024Stated by the company or an official page
Redwood1X TechnologiesTeleoperation data1X trained its Redwood model on a large dataset of teleoperated and autonomous episodes from its EVE and NEO robots, gathered in 1X offices and employee homes.10 Jun 2025Stated by the company or an official page
RFM-1MIT Technology ReviewData mixMIT Technology Review reported that RFM-1 was trained on years of data from Covariant's warehouse item-picking robots plus words and videos from the internet.11 Mar 2024Reported by an outlet
RFM-1IEEE SpectrumEpisodesIEEE Spectrum reported that Covariant trained the 8-billion-parameter RFM-1 on tens of millions of trajectories from its warehouse fleet; cofounder Pieter Abbeel said the fleet yields a million trajectories every few weeks.11 Mar 2024Reported by an outlet
RoboCatGoogle DeepMindTeleoperation dataTo teach RoboCat a new task or robot, DeepMind collected 100 to 1,000 demonstrations with a human-controlled robot arm, then let RoboCat generate more training data itself.20 Jun 2023Stated by the company or an official page
RT-2Google DeepMindData mixRT-2 was co-fine-tuned on web data and on RT-1 robot demonstration data collected with 13 robots over 17 months in an office kitchen.28 Jul 2023Stated by the company or an official page
RT-X (Open X-Embodiment)arXiv (Open X-Embodiment Collaboration)Robot typesThe Open X-Embodiment dataset behind the RT-X models pools data from 22 different robots, collected with 21 institutions, covering 527 skills.3 Oct 2023Stated by the company or an official page
RT-X (Open X-Embodiment)Google DeepMindEpisodesGoogle DeepMind said the Open X-Embodiment dataset used to train RT-1-X and RT-2-X spans more than 1 million episodes, 500 skills and 150,000 tasks.3 Oct 2023Stated by the company or an official page
Skild BrainSkild AIHuman videoSkild AI said it fine-tunes its model to perform new skills by watching human videos, plus a small amount of robot data that the post puts at under one hour.12 Jan 2026Stated by the company or an official page
Skild BrainSkild AIRobot typesSkild AI said it trained its omni-bodied brain to control 100,000 different simulated robots, over a millennium of simulated time.24 Sep 2025Stated by the company or an official page
Skild BrainSkild AIData mixSkild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.29 Jul 2025Stated by the company or an official page
Skild S1Skild AIHours of training dataFor S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours.18 Aug 2026Stated by the company or an official page
SmolVLAarXiv (Hugging Face)EpisodesHugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets.3 Jun 2025Stated by the company or an official page
SmolVLAHugging FaceData mixSmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all.3 Jun 2025Stated by the company or an official page
V-JEPA 2arXiv (Meta FAIR)Human videoMeta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.11 Jun 2025Stated by the company or an official page
V-JEPA 2-ACarXiv (Meta FAIR)Hours of training dataMeta post-trained the action-conditioned V-JEPA 2-AC world model on less than 62 hours of unlabeled robot video from the DROID dataset.11 Jun 2025Stated by the company or an official page
Xiaomi-Robotics-1arXiv (Xiaomi Robotics Team)Hours of training dataXiaomi's robotics team pre-trained its Xiaomi-Robotics-1 vision-language-action model on over 100,000 hours of real-world manipulation trajectories collected with UMI devices.16 Jul 2026Stated by the company or an official page
λ0 (HumanVerse)arXiv (Lambda-humanoid Team)Human videoThe Lambda-humanoid Team introduced HumanVerse, a 500-hour dataset of open-world human loco-manipulation pairing egocentric video with synchronized body and hand motion, to train its λ0 humanoid model.13 Oct 2026Stated by the company or an official page
π*0.6Physical IntelligenceTeleoperation dataPhysical Intelligence says robot foundation models today mainly learn from demonstrations people collect by hand, such as teleoperation; for π*0.6 it adds expert teleoperator corrections and the robot's own autonomous experience.17 Nov 2025Stated by the company or an official page
π0Physical IntelligenceReleasePhysical Intelligence released the code and weights of π0 as open source on February 4, 2025, saying 1 to 20 hours of data had been enough to fine-tune it to a variety of tasks.4 Feb 2025Stated by the company or an official page
π0arXiv (Physical Intelligence)Hours of training dataPhysical Intelligence pre-trained π0 on about 10,000 hours of its own dexterous robot manipulation data, alongside open datasets such as Open X-Embodiment.31 Oct 2024Stated by the company or an official page
π0arXiv (Physical Intelligence)Robot typesPhysical Intelligence's own π0 pre-training data covers 7 robot configurations and 68 tasks (its launch blog describes the data as coming from 8 distinct robots); the mixture also includes the Open X-Embodiment data from 22 robots.31 Oct 2024Stated by the company or an official page
π0.5arXiv (Physical Intelligence)Hours of training dataπ0.5 uses about 400 hours of data from mobile manipulators doing household tasks in about 100 different homes, the slice of its training data closest to its test tasks.22 Apr 2025Stated by the company or an official page
π0.5arXiv (Physical Intelligence)Data mix97.6% of the training examples in π0.5's first training phase do not come from mobile manipulators doing household tasks but from other robots and web data.22 Apr 2025Stated by the company or an official page
π0.5 (human-to-robot transfer)Physical IntelligenceHuman videoPhysical Intelligence found that egocentric video of humans, recorded with wearable cameras, roughly doubled performance on tasks where robot data is limited once its models were trained on enough robot data.16 Dec 2025Stated by the company or an official page
π0.7Physical IntelligenceData mixπ0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.16 Apr 2026Stated by the company or an official page

Base: All 79 facts. "Reported" means an outlet reports it from sources the company has not confirmed.

Amounts stay in the currency and form their source gives; nothing is converted or totalled.

Credit: compiled by RobotTraining.jobs Research, How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition.

Robotics labs · Fleet operators · Data vendors · Research teams

Buy the data behind this report

This report sums up records we gather ourselves, day after day:

  • 1,432 robot training roles followed, history included

    Every role on 100 employer job boards, read each day since 12 Sep 2026: its pay, its place, and the first and last day it was up.

  • 536 apply clicks over 28 days

    Which robot training roles people really apply to, by kind of work, pay and place. The clicks happen here, so no one else can count them.

  • 25 sourced industry datasets

    Funding, revenue, deals, fleets, deployments and prices across the robot training industry, each record linked to the publication it comes from.

On offer

  • 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.

Need the people behind this data for your own robots? (you contract them yourself) or (we deliver the robot data).

Investing in robotics, AV or robot-data companies? The Robot hiring radar is these records role by role, every day, as files.

Robot hiring brief: free, once a month

The robot-training job market in figures, on the first weekday of each month: open roles, what opened and closed, who added and who cut, and posted pay by kind of work. To your work email.

How it was measured

Developers' papers, model cards and announcements, and reporting by major outlets where a developer has not published, each linked 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.

Foundation model
A large model trained on data from many tasks and robots, then adapted to new ones; vision-language-action (VLA) models are the common kind.
Teleoperation data
Recordings of people steering robots remotely or in VR, the work this site's teleoperation listings pay for.

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.
  • Developers count data differently (robot hours, human video hours, episodes, simulated trajectories), so figures are not compared one to one.
  • Many developers disclose little about their data; a missing figure means it is not public.

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
  • October 2026 edition (you are here)Will lock when the next daily refresh runs

Quoting this report

You may quote, chart or republish these figures under CC BY 4.0, provided you link back to this page. This edition is locked: its numbers will not change.

Citation

RobotTraining.jobs (2026). How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition. https://robottraining.jobs/research/robot-foundation-model-training-data/2026-10

BibTeX

@misc{robottrainingjobs_robot_foundation_model_training_data_2026_10,
  author = {{RobotTraining.jobs}},
  title = {How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, October 2026 edition},
  year = {2026},
  howpublished = {\url{https://robottraining.jobs/research/robot-foundation-model-training-data/2026-10}},
}
Files: every figure in one CSV (a row per value) or as JSON, emailed to your work address.Other open data: datasets and API

Keep reading

Browse all 91 reports