What the data shows
- 1
GR00T N1 is the most documented of the 42 robot models (4 sourced facts).
- 2
GEN-1 (Generalist AI) states the most hours of training data on record: 500,000 hours.
- 3
44 statements from 32 models say how much of their data came from teleoperation, human video or simulation.
- 4
The most recent fact on record (18 Aug 2026): For S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours.
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?
| Model | Training data as stated | Sourced facts |
|---|---|---|
| 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. | 4 |
| 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. | 4 |
| 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. | 3 |
| GEN-1Generalist AI | Generalist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data. | 3 |
| GO-1 (Genie Operator-1)arXiv (AgiBot) | The 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. | 3 |
| 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. | 3 |
| 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. | 3 |
| Skild BrainSkild AI | Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data. | 3 |
| 1X World Model (1XWM)1X Technologies | 1X 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. | 2 |
| ACT-1Sunday Robotics | Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate. | 2 |
| Dyna-2Dyna Robotics | n/a | 2 |
| GEN-0Generalist AI | Generalist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week. | 2 |
| 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. | 2 |
| OptimusBusiness Insider (via AOL) | n/a | 2 |
| 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. | 2 |
| RFM-1MIT Technology Review | MIT 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. | 2 |
| RT-X (Open X-Embodiment)Google DeepMind | Google 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. | 2 |
| SmolVLAarXiv (Hugging Face) | Hugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets. | 2 |
| ACT-2Sunday Robotics | Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware. | 1 |
| Atlas Large Behavior ModelBoston Dynamics | n/a | 1 |
| Diffusion Policy behavior modelToyota USA Newsroom | n/a | 1 |
| Gemini RoboticsarXiv (Google DeepMind) | n/a | 1 |
| Gemini Robotics 1.5arXiv (Google DeepMind) | n/a | 1 |
| Gemini Robotics On-DeviceGoogle DeepMind | n/a | 1 |
| GR-2arXiv (ByteDance Research) | n/a | 1 |
| GR00T N1.6NVIDIA Research (GEAR) | n/a | 1 |
| 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. | 1 |
| HelixFigure AI | n/a | 1 |
| Helix 02Figure AI | Helix 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. | 1 |
| Helix (Project Go-Big)Figure AI | n/a | 1 |
| OctoarXiv | Octo was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection. | 1 |
| OpenVLAarXiv | OpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset. | 1 |
| π*0.6Physical Intelligence | n/a | 1 |
| π0.5 (human-to-robot transfer)Physical Intelligence | n/a | 1 |
| π0.7Physical Intelligence | π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies. | 1 |
| RDT-1BRDT project (Tsinghua University) | RDT-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. | 1 |
| Redwood1X Technologies | n/a | 1 |
| RoboCatGoogle DeepMind | n/a | 1 |
| RT-2Google DeepMind | RT-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. | 1 |
| Skild S1Skild AI | For S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours. | 1 |
| V-JEPA 2arXiv (Meta FAIR) | n/a | 1 |
| 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. | 1 |
Base: 42 robot models, 70 facts, each linked to the page that states it (reviewed 18 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, checked 18 Oct 2026.
Figure 2
Models by stated hours of training data
Which robot models were trained on the most hours of data?
| Model | Hours |
|---|---|
| GEN-1Generalist AI | 500,000 |
| GEN-0Generalist AI | 270,000 |
| π0arXiv (Physical Intelligence) | 10,000 |
| Large Behavior Model (LBM)arXiv (Toyota Research Institute) | 1,700 |
| Helix 02Figure AI | 1,000 |
| G0arXiv (Galaxea) | 500 |
| π0.5arXiv (Physical Intelligence) | 400 |
| 1X World Model (1XWM)1X Technologies | 70 |
| V-JEPA 2-ACarXiv (Meta FAIR) | 62 |
Base: 9 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, checked 18 Oct 2026.
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?
| Model | As stated |
|---|---|
| Dyna-2Dyna Robotics | Dyna 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. |
| Dyna-2Dyna Robotics | Dyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands. |
| ACT-2Sunday Robotics | Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware. |
| π0.7Physical Intelligence | π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies. |
| GEN-1Generalist AI | 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. |
| 1X World Model (1XWM)1X Technologies | 1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video. |
| Skild BrainSkild AI | 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. |
| π0.5 (human-to-robot transfer)Physical Intelligence | Physical 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. |
| 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. |
| ACT-1Sunday Robotics | Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove. |
| ACT-1Sunday Robotics | Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate. |
| π*0.6Physical Intelligence | Physical 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. |
| GEN-0Generalist AI | Generalist 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. |
| Helix (Project Go-Big)Figure AI | Figure 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| Atlas Large Behavior ModelBoston Dynamics | Boston Dynamics collects the training data for Atlas's large behavior model with a VR-based teleoperation system, on the real robot and in simulation. |
| Skild BrainSkild AI | Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data. |
| 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. |
| GR-3arXiv (ByteDance Seed) | For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total. |
| 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| Redwood1X Technologies | 1X 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. |
| SmolVLAHugging Face | SmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all. |
| GR00T N1.5NVIDIA Newsroom | NVIDIA 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| 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. |
| HelixFigure AI | Figure 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. |
| GR-2arXiv (ByteDance Research) | ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories. |
| RFM-1MIT Technology Review | MIT 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. |
| Diffusion Policy behavior modelToyota USA Newsroom | TRI 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. |
| RT-2Google DeepMind | RT-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. |
| RoboCatGoogle DeepMind | To 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. |
Base: 44 facts from 32 robot models.
Credit: compiled by RobotTraining.jobs Research, How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video, checked 18 Oct 2026.
Figure 4
Every sourced fact
What exactly has been published about each model?
| Model | Kind | Fact |
|---|---|---|
| 1X World Model (1XWM)1X Technologies | Human video | 1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video. |
| 1X World Model (1XWM)1X Technologies | Hours of training data | 1X 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. |
| ACT-1Sunday Robotics | Teleoperation data | Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove. |
| ACT-1Sunday Robotics | Data mix | Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate. |
| ACT-2Sunday Robotics | Data mix | Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware. |
| Atlas Large Behavior ModelBoston Dynamics | Teleoperation data | Boston Dynamics collects the training data for Atlas's large behavior model with a VR-based teleoperation system, on the real robot and in simulation. |
| Diffusion Policy behavior modelToyota USA Newsroom | Teleoperation data | TRI 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. |
| Dyna-2Dyna Robotics | Human video | Dyna 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. |
| Dyna-2Dyna Robotics | Teleoperation data | Dyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands. |
| G0arXiv (Galaxea) | Hours of training data | 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. |
| G0arXiv (Galaxea) | Teleoperation data | 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. |
| G0arXiv (Galaxea) | Data mix | 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. |
| Gemini RoboticsarXiv (Google DeepMind) | Teleoperation data | 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. |
| Gemini Robotics 1.5arXiv (Google DeepMind) | Robot types | Gemini 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. |
| Gemini Robotics On-DeviceGoogle DeepMind | Robot types | Gemini 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. |
| GEN-0Generalist AI | Hours of training data | Generalist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week. |
| GEN-0Generalist AI | Data mix | Generalist 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. |
| GEN-1Generalist AI | Hours of training data | Generalist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data. |
| GEN-1Generalist AI | Human video | 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. |
| GEN-1Generalist AI | Hours of training data | Each GEN-1 task result Generalist showed was built with about one hour of robot data. |
| GO-1 (Genie Operator-1)arXiv (AgiBot) | Episodes | The 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. |
| GO-1 (Genie Operator-1)arXiv (AgiBot) | Teleoperation data | AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation. |
| GO-1 (Genie Operator-1)Newsfile Corp (AgiBot release) | Data mix | 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. |
| GR-2arXiv (ByteDance Research) | Human video | ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories. |
| GR-3arXiv (ByteDance Seed) | Data mix | ByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning. |
| GR-3arXiv (ByteDance Seed) | Teleoperation data | For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total. |
| GR-3arXiv (ByteDance Seed) | Data mix | ByteDance said human trajectories recorded with VR devices come in at about 450 per hour, against about 250 per hour for teleoperated robot trajectories. |
| GR00T N1arXiv (NVIDIA) | Teleoperation data | 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. |
| GR00T N1arXiv (NVIDIA) | Data mix | 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. |
| GR00T N1arXiv (NVIDIA) | Episodes | NVIDIA 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. |
| GR00T N1arXiv (NVIDIA) | Human video | 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. |
| GR00T N1.5NVIDIA Research (GEAR) | Data mix | 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. |
| GR00T N1.5NVIDIA Newsroom | Data mix | NVIDIA 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. |
| GR00T N1.6NVIDIA Research (GEAR) | Teleoperation data | 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. |
| GraspVLAarXiv (Galbot, Peking University) | Data mix | Galbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation. |
| HelixFigure AI | Teleoperation data | Figure 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. |
| Helix (Project Go-Big)Figure AI | Human video | Figure 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. |
| Helix 02Figure AI | Hours of training data | Helix 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. |
| Large Behavior Model (LBM)arXiv (Toyota Research Institute) | Hours of training data | TRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data. |
| Large Behavior Model (LBM)arXiv (Toyota Research Institute) | Episodes | TRI'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. |
| Large Behavior Model (LBM)arXiv (Toyota Research Institute) | Teleoperation 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. |
| Large Behavior Model (LBM)arXiv (Toyota Research Institute) | Data mix | 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. |
| OctoarXiv | Episodes | Octo was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection. |
| OpenVLAarXiv | Episodes | OpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset. |
| OptimusBusiness Insider (via AOL) | Human video | 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. |
| OptimusBusiness Insider (via AOL) | Teleoperation data | 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. |
| RDT-1BRDT project (Tsinghua University) | Episodes | RDT-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. |
| Redwood1X Technologies | Teleoperation data | 1X 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. |
| RFM-1MIT Technology Review | Data mix | MIT 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. |
| RFM-1IEEE Spectrum | Episodes | IEEE 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. |
| RoboCatGoogle DeepMind | Teleoperation data | To 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. |
| RT-2Google DeepMind | Data mix | RT-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. |
| RT-X (Open X-Embodiment)arXiv (Open X-Embodiment Collaboration) | Robot types | The Open X-Embodiment dataset behind the RT-X models pools data from 22 different robots, collected with 21 institutions, covering 527 skills. |
| RT-X (Open X-Embodiment)Google DeepMind | Episodes | Google 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. |
| Skild BrainSkild AI | Human video | 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. |
| Skild BrainSkild AI | Robot types | Skild AI said it trained its omni-bodied brain to control 100,000 different simulated robots, over a millennium of simulated time. |
| Skild BrainSkild AI | Data mix | Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data. |
| Skild S1Skild AI | Hours of training data | For S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours. |
| SmolVLAarXiv (Hugging Face) | Episodes | Hugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets. |
| SmolVLAHugging Face | Data mix | SmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all. |
| V-JEPA 2arXiv (Meta FAIR) | Human video | Meta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video. |
| V-JEPA 2-ACarXiv (Meta FAIR) | Hours of training data | 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. |
| π*0.6Physical Intelligence | Teleoperation data | Physical 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. |
| π0Physical Intelligence | Release | Physical 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. |
| π0arXiv (Physical Intelligence) | Hours of training data | 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. |
| π0arXiv (Physical Intelligence) | Robot types | Physical 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. |
| π0.5arXiv (Physical Intelligence) | Hours of training data | 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. |
| π0.5arXiv (Physical Intelligence) | Data mix | 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. |
| π0.5 (human-to-robot transfer)Physical Intelligence | Human video | Physical 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. |
| π0.7Physical Intelligence | Data mix | π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies. |
Base: All 70 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, 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
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
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 Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video (live data, 20 Jun 2023 to 18 Oct 2026). https://robottraining.jobs/research/robot-foundation-model-training-data (accessed 18 Oct 2026)
BibTeX
@misc{robottrainingjobs_robot_foundation_model_training_data_live,
author = {{RobotTraining.jobs}},
title = {How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video (live data, 20 Jun 2023 to 18 Oct 2026)},
year = {2026},
howpublished = {\url{https://robottraining.jobs/research/robot-foundation-model-training-data}},
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
- How Robotics Labs Collect Their Own Training Data: Facilities, Teams, Hours and PayWhich robotics, humanoid, embodied-AI, AV and drone companies collect training data with their own people, how (teleoperation, wearable cameras, safety drivers), at what scale, and what do they pay?Open report
- What Robot Training Data Costs: Buyer Prices and What Recorders Are PaidWhat does an hour of robot, teleoperation or first-person training data cost the company that buys it, and what are the people who record it or teleoperate paid?Open report
- Open Robot Training Datasets: Size, Collection Method and LicenceWhich public datasets teach robots, embodied AI, self-driving systems and drones, who released them, how many hours and episodes do they hold and how was the data collected?Open report
- Teleoperation Companies: Remote Driving and Remote Robot Operators, Pay and FundingWhich companies run remote driving or remote robot operation, where do their operators work, what are they paid, who are their customers and how are they funded?Open report
- Who Watches the Robotaxis: Remote Assistance Teams, Ratios, Shifts and TrainingHow many people remotely assist robotaxis, autonomous trucks and delivery robots, where do they work, what shifts do they keep, what training do they need and how many vehicles does each one watch?Open report
- Humanoid Robots in the Real World: Units Shipped, Deployments, Customers and PricesHow many humanoid robots have been built and shipped, where are they deployed, who has ordered them and what do they cost?Open report
- Robotaxi Fleets and Cities: Fleet Sizes, Paid Rides, Safety Operators and PermitsHow many vehicles does each robotaxi service run, in which cities, how many paid rides does it give, and does a safety driver or monitor still ride along?Open report
- Drone Delivery at Scale: Deliveries, Markets, FAA Approvals and Drones per PilotHow many deliveries have Zipline, Wing, Amazon Prime Air, Manna and others made, where do they fly, under which approvals, and how many drones does one remote pilot oversee?Open report
- Apps and Programs That Pay You to Record Robot Training DataWhich apps and programs pay people to film household chores or work tasks, wear a head camera or teleoperate from home for robot training, what do they pay, who can join, which device is needed and how do they pay out?Open report
- The Hardware Behind Robot Training Data: Headsets, Capture Glasses, Grippers, Leader Arms and SuitsWhich VR headsets, first-person cameras, handheld grippers, leader-follower arms, gloves and motion-capture rigs are used to collect robot training data, what do they cost and who uses them?Open report
- Autonomous Trucking in Operation: Driverless Lanes, Miles, Observers and CustomersWhich autonomous trucks run without a driver, on which lanes, for which freight customers, how many miles have they driven, and does a safety driver or observer still ride in the cab?Open report
- China's Robot Data Collection Centres: Size, Robots, Collectors, Pay and Data ProducedWhere are China's state-backed and company-run embodied-AI data collection and training centres, when did they open, how big are they, how many robots and data collectors work there, what are collectors paid and how much data have they produced?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