{"name":"How Much Data Robot Foundation Models Are Trained On: Hours, Teleoperation and Human Video (live data, 20 Jun 2023 to 18 Oct 2026)","question":"How many hours and episodes of data were robot foundation models (pi0, GR00T, Gemini Robotics, Helix, GEN-0 and others) trained on, and how much of it came from people teleoperating robots or filming their own hands?","source":"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.","publisher":"RobotTraining.jobs","url":"https://robottraining.jobs/research/robot-foundation-model-training-data","license":"https://creativecommons.org/licenses/by/4.0/","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)","frozen_at":null,"report":"robot-foundation-model-training-data","method":1,"edition":null,"window":{"from":"2023-06-20","to":"2026-10-18","days":1217},"computedAt":"2026-10-18T20:07:26.217Z","summary":"42 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; 70 facts, each linked to its source.","stats":[{"id":"subjects","value":"42","label":"robot models","figure":"subjects"},{"id":"facts","value":"70","label":"sourced facts","figure":"facts"}],"findings":[{"id":"most-documented","figure":"subjects","text":"GR00T N1 is the most documented of the 42 robot models (4 sourced facts)."},{"id":"hours","figure":"hours","text":"GEN-1 (Generalist AI) states the most hours of training data on record: 500,000 hours."},{"id":"teleop","figure":"teleop","text":"44 statements from 32 models say how much of their data came from teleoperation, human video or simulation."},{"id":"latest","figure":"facts","text":"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."}],"figures":[{"id":"subjects","title":"Robot foundation models and the data behind them","question":"How much human-made data was each robot foundation model trained on, and where did it come from?","columns":[{"key":"subject","label":"Model","unit":"text"},{"key":"company","label":"Company","unit":"text","secondary":true},{"key":"data","label":"Training data as stated","unit":"text","secondary":false},{"key":"teleop","label":"Teleoperation or human video","unit":"text","secondary":true},{"key":"release","label":"Release","unit":"text","secondary":true},{"key":"facts","label":"Sourced facts","unit":"count","bar":true}],"rows":[{"key":"gr00t-n1","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","company":"NVIDIA","facts":4,"data":"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.","teleop":"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.","release":"n/a"}},{"key":"tri-lbm","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","company":"Toyota Research Institute","facts":4,"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.","teleop":"The real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.","release":"n/a"}},{"key":"g0","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","company":"Galaxea","facts":3,"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.","teleop":"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.","release":"n/a"}},{"key":"gen-1","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","company":"Generalist AI","facts":3,"data":"Generalist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.","teleop":"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.","release":"n/a"}},{"key":"go-1","source":{"label":"arXiv (AgiBot)","url":"https://arxiv.org/html/2503.06669v4","date":"2025-08-04"},"cells":{"subject":"GO-1 (Genie Operator-1)","company":"AgiBot","facts":3,"data":"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.","teleop":"AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.","release":"n/a"}},{"key":"gr-3","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/abs/2507.15493","date":"2025-07-21"},"cells":{"subject":"GR-3","company":"ByteDance Seed","facts":3,"data":"ByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.","teleop":"For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.","release":"n/a"}},{"key":"pi0","href":"/platforms/physical-intelligence","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2410.24164v1","date":"2024-10-31"},"cells":{"subject":"π0","company":"Physical Intelligence","facts":3,"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.","teleop":"n/a","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."}},{"key":"skild-brain","href":"/platforms/skild-ai","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain","date":"2025-07-29"},"cells":{"subject":"Skild Brain","company":"Skild AI","facts":3,"data":"Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.","teleop":"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.","release":"n/a"}},{"key":"1x-world-model","href":"/platforms/1x","source":{"label":"1X Technologies","url":"https://1x.tech/discover/world-model-self-learning","date":"2026-01-12"},"cells":{"subject":"1X World Model (1XWM)","company":"1X Technologies","facts":2,"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.","teleop":"1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.","release":"n/a"}},{"key":"act-1","href":"/platforms/sunday-robotics","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/no-robot-data","date":"2025-11-19"},"cells":{"subject":"ACT-1","company":"Sunday Robotics","facts":2,"data":"Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.","teleop":"Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.","release":"n/a"}},{"key":"dyna-2","href":"/platforms/dyna-robotics","source":{"label":"Dyna Robotics","url":"https://www.dyna.co/dyna-2","date":"2026-08-01"},"cells":{"subject":"Dyna-2","company":"Dyna Robotics","facts":2,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"gen-0","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/nov-04-2025-GEN-0","date":"2025-11-04"},"cells":{"subject":"GEN-0","company":"Generalist AI","facts":2,"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.","teleop":"n/a","release":"n/a"}},{"key":"gr00t-n1-5","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_5/","date":"2025-06-11"},"cells":{"subject":"GR00T N1.5","company":"NVIDIA","facts":2,"data":"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.","teleop":"n/a","release":"n/a"}},{"key":"optimus","href":"/platforms/tesla-optimus","source":{"label":"Business Insider (via AOL)","url":"https://www.aol.com/tesla-shifted-optimus-training-strategy-092502010.html","date":"2025-08-25"},"cells":{"subject":"Optimus","company":"Tesla","facts":2,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"pi0-5","href":"/platforms/physical-intelligence","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2504.16054v1","date":"2025-04-22"},"cells":{"subject":"π0.5","company":"Physical Intelligence","facts":2,"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.","teleop":"n/a","release":"n/a"}},{"key":"rfm-1","source":{"label":"MIT Technology Review","url":"https://www.technologyreview.com/2024/03/11/1089653/an-openai-spinoff-has-built-an-ai-model-that-helps-robots-learn-tasks-like-humans/","date":"2024-03-11"},"cells":{"subject":"RFM-1","company":"Covariant","facts":2,"data":"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.","teleop":"n/a","release":"n/a"}},{"key":"rt-x","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/scaling-up-learning-across-many-different-robot-types/","date":"2023-10-03"},"cells":{"subject":"RT-X (Open X-Embodiment)","company":"Google DeepMind","facts":2,"data":"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.","teleop":"n/a","release":"n/a"}},{"key":"smolvla","source":{"label":"arXiv (Hugging Face)","url":"https://arxiv.org/html/2506.01844v1","date":"2025-06-02"},"cells":{"subject":"SmolVLA","company":"Hugging Face","facts":2,"data":"Hugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets.","teleop":"n/a","release":"n/a"}},{"key":"act-2","href":"/platforms/sunday-robotics","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/act-2-preview","date":"2026-07-17"},"cells":{"subject":"ACT-2","company":"Sunday Robotics","facts":1,"data":"Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.","teleop":"n/a","release":"n/a"}},{"key":"atlas-large-behavior-model","href":"/platforms/boston-dynamics","source":{"label":"Boston Dynamics","url":"https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/","date":"2025-08-20"},"cells":{"subject":"Atlas Large Behavior Model","company":"Boston Dynamics and Toyota Research Institute","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"diffusion-policy-behavior-model","source":{"label":"Toyota USA Newsroom","url":"https://pressroom.toyota.com/toyota-research-institute-unveils-breakthrough-in-teaching-robots-new-behaviors/","date":"2023-09-19"},"cells":{"subject":"Diffusion Policy behavior model","company":"Toyota Research Institute","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"gemini-robotics","source":{"label":"arXiv (Google DeepMind)","url":"https://arxiv.org/html/2503.20020v1","date":"2025-03-25"},"cells":{"subject":"Gemini Robotics","company":"Google DeepMind","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"gemini-robotics-1-5","source":{"label":"arXiv (Google DeepMind)","url":"https://arxiv.org/html/2510.03342v1","date":"2025-10-02"},"cells":{"subject":"Gemini Robotics 1.5","company":"Google DeepMind","facts":1,"data":"n/a","teleop":"n/a","release":"n/a"}},{"key":"gemini-robotics-on-device","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/","date":"2025-06-24"},"cells":{"subject":"Gemini Robotics On-Device","company":"Google DeepMind","facts":1,"data":"n/a","teleop":"n/a","release":"n/a"}},{"key":"gr-2","source":{"label":"arXiv (ByteDance Research)","url":"https://arxiv.org/abs/2410.06158","date":"2024-10-08"},"cells":{"subject":"GR-2","company":"ByteDance Seed","facts":1,"data":"n/a","teleop":"ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.","release":"n/a"}},{"key":"gr00t-n1-6","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_6/","date":"2025-12-15"},"cells":{"subject":"GR00T N1.6","company":"NVIDIA","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"graspvla","source":{"label":"arXiv (Galbot, Peking University)","url":"https://arxiv.org/abs/2505.03233","date":"2025-05-06"},"cells":{"subject":"GraspVLA","company":"Galbot","facts":1,"data":"Galbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.","teleop":"n/a","release":"n/a"}},{"key":"helix","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix","date":"2025-02-20"},"cells":{"subject":"Helix","company":"Figure AI","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"helix-02","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix-02","date":"2026-01-27"},"cells":{"subject":"Helix 02","company":"Figure AI","facts":1,"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.","teleop":"n/a","release":"n/a"}},{"key":"helix-project-go-big","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/project-go-big","date":"2025-09-18"},"cells":{"subject":"Helix (Project Go-Big)","company":"Figure AI","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"octo","source":{"label":"arXiv","url":"https://arxiv.org/abs/2405.12213","date":"2024-05-20"},"cells":{"subject":"Octo","company":"UC Berkeley, Stanford, CMU and Google DeepMind (Octo Model Team)","facts":1,"data":"Octo was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection.","teleop":"n/a","release":"n/a"}},{"key":"openvla","source":{"label":"arXiv","url":"https://arxiv.org/abs/2406.09246","date":"2024-06-13"},"cells":{"subject":"OpenVLA","company":"Stanford, UC Berkeley, Toyota Research Institute, Google DeepMind and others","facts":1,"data":"OpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset.","teleop":"n/a","release":"n/a"}},{"key":"pi-star-0-6","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pistar06","date":"2025-11-17"},"cells":{"subject":"π*0.6","company":"Physical Intelligence","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"pi0-5-human-to-robot-transfer","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/research/human_to_robot","date":"2025-12-16"},"cells":{"subject":"π0.5 (human-to-robot transfer)","company":"Physical Intelligence","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"pi0-7","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pi07","date":"2026-04-16"},"cells":{"subject":"π0.7","company":"Physical Intelligence","facts":1,"data":"π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.","teleop":"n/a","release":"n/a"}},{"key":"rdt-1b","source":{"label":"RDT project (Tsinghua University)","url":"https://rdt-robotics.github.io/rdt-robotics/","date":"2026-10-18"},"cells":{"subject":"RDT-1B","company":"Tsinghua University","facts":1,"data":"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.","teleop":"n/a","release":"n/a"}},{"key":"redwood","href":"/platforms/1x","source":{"label":"1X Technologies","url":"https://www.1x.tech/discover/redwood-ai","date":"2025-06-10"},"cells":{"subject":"Redwood","company":"1X Technologies","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"robocat","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/robocat-a-self-improving-robotic-agent/","date":"2023-06-20"},"cells":{"subject":"RoboCat","company":"Google DeepMind","facts":1,"data":"n/a","teleop":"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.","release":"n/a"}},{"key":"rt-2","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action/","date":"2023-07-28"},"cells":{"subject":"RT-2","company":"Google DeepMind","facts":1,"data":"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.","teleop":"n/a","release":"n/a"}},{"key":"skild-s1","href":"/platforms/skild-ai","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/s1","date":"2026-08-18"},"cells":{"subject":"Skild S1","company":"Skild AI","facts":1,"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.","teleop":"n/a","release":"n/a"}},{"key":"v-jepa-2","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2","company":"Meta","facts":1,"data":"n/a","teleop":"Meta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.","release":"n/a"}},{"key":"v-jepa-2-ac","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2-AC","company":"Meta","facts":1,"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.","teleop":"n/a","release":"n/a"}}],"sample":"42 robot models, 70 facts, each linked to the page that states it (reviewed 18 Oct 2026).","notes":["A model appears once its developer or a major outlet has said something about its training data."]},{"id":"hours","title":"Models by stated hours of training data","question":"Which robot models were trained on the most hours of data?","columns":[{"key":"subject","label":"Model","unit":"text"},{"key":"number","label":"Hours","unit":"count","bar":true},{"key":"fact","label":"As stated","unit":"text","secondary":true},{"key":"date","label":"Date","unit":"text","secondary":true}],"rows":[{"key":"gen-1-hours","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","number":500000,"fact":"Generalist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.","date":"2 Apr 2026"}},{"key":"gen-0-hours","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/nov-04-2025-GEN-0","date":"2025-11-04"},"cells":{"subject":"GEN-0","number":270000,"fact":"Generalist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week.","date":"4 Nov 2025"}},{"key":"pi0-hours","href":"/platforms/physical-intelligence","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2410.24164v1","date":"2024-10-31"},"cells":{"subject":"π0","number":10000,"fact":"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.","date":"31 Oct 2024"}},{"key":"tri-lbm-hours","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","number":1700,"fact":"TRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data.","date":"7 Jul 2025"}},{"key":"helix-02-hours","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix-02","date":"2026-01-27"},"cells":{"subject":"Helix 02","number":1000,"fact":"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.","date":"27 Jan 2026"}},{"key":"g0-hours","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","number":500,"fact":"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.","date":"30 Aug 2025"}},{"key":"pi0-5-hours","href":"/platforms/physical-intelligence","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2504.16054v1","date":"2025-04-22"},"cells":{"subject":"π0.5","number":400,"fact":"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.","date":"22 Apr 2025"}},{"key":"1x-world-model-hours","href":"/platforms/1x","source":{"label":"1X Technologies","url":"https://1x.tech/discover/world-model-self-learning","date":"2026-01-12"},"cells":{"subject":"1X World Model (1XWM)","number":70,"fact":"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.","date":"12 Jan 2026"}},{"key":"v-jepa-2-ac-hours","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2-AC","number":62,"fact":"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.","date":"11 Jun 2025"}}],"sample":"9 robot models with a stated figure; each row is the largest one on record for it.","notes":["Hours as each developer counts them (robot data, human video or both); they are not comparable one to one."]},{"id":"teleop","title":"How much of it people made by teleoperating robots or filming themselves","question":"What share of robot model training data comes from teleoperation and from human video?","columns":[{"key":"subject","label":"Model","unit":"text"},{"key":"value","label":"As stated","unit":"text"},{"key":"date","label":"Date","unit":"text","secondary":true}],"rows":[{"key":"dyna-2-teleop-0","href":"/platforms/dyna-robotics","source":{"label":"Dyna Robotics","url":"https://www.dyna.co/dyna-2","date":"2026-08-01"},"cells":{"subject":"Dyna-2","value":"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.","date":"1 Aug 2026"}},{"key":"dyna-2-teleop-1","href":"/platforms/dyna-robotics","source":{"label":"Dyna Robotics","url":"https://www.dyna.co/dyna-2","date":"2026-08-01"},"cells":{"subject":"Dyna-2","value":"Dyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands.","date":"1 Aug 2026"}},{"key":"act-2-teleop-2","href":"/platforms/sunday-robotics","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/act-2-preview","date":"2026-07-17"},"cells":{"subject":"ACT-2","value":"Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.","date":"17 Jul 2026"}},{"key":"pi0-7-teleop-3","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pi07","date":"2026-04-16"},"cells":{"subject":"π0.7","value":"π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.","date":"16 Apr 2026"}},{"key":"gen-1-teleop-4","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","value":"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.","date":"2 Apr 2026"}},{"key":"1x-world-model-teleop-5","href":"/platforms/1x","source":{"label":"1X Technologies","url":"https://1x.tech/discover/world-model-self-learning","date":"2026-01-12"},"cells":{"subject":"1X World Model (1XWM)","value":"1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.","date":"12 Jan 2026"}},{"key":"skild-brain-teleop-6","href":"/platforms/skild-ai","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/learning-by-watching","date":"2026-01-12"},"cells":{"subject":"Skild Brain","value":"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.","date":"12 Jan 2026"}},{"key":"pi0-5-human-to-robot-transfer-teleop-7","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/research/human_to_robot","date":"2025-12-16"},"cells":{"subject":"π0.5 (human-to-robot transfer)","value":"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.","date":"16 Dec 2025"}},{"key":"gr00t-n1-6-teleop-8","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_6/","date":"2025-12-15"},"cells":{"subject":"GR00T N1.6","value":"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.","date":"15 Dec 2025"}},{"key":"act-1-teleop-9","href":"/platforms/sunday-robotics","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/no-robot-data","date":"2025-11-19"},"cells":{"subject":"ACT-1","value":"Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.","date":"19 Nov 2025"}},{"key":"act-1-teleop-10","href":"/platforms/sunday-robotics","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/no-robot-data","date":"2025-11-19"},"cells":{"subject":"ACT-1","value":"Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.","date":"19 Nov 2025"}},{"key":"pi-star-0-6-teleop-11","href":"/platforms/physical-intelligence","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pistar06","date":"2025-11-17"},"cells":{"subject":"π*0.6","value":"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.","date":"17 Nov 2025"}},{"key":"gen-0-teleop-12","href":"/platforms/generalist-ai","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/nov-04-2025-GEN-0","date":"2025-11-04"},"cells":{"subject":"GEN-0","value":"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.","date":"4 Nov 2025"}},{"key":"helix-project-go-big-teleop-13","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/project-go-big","date":"2025-09-18"},"cells":{"subject":"Helix (Project Go-Big)","value":"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.","date":"18 Sep 2025"}},{"key":"g0-teleop-14","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","value":"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.","date":"30 Aug 2025"}},{"key":"g0-teleop-15","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","value":"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.","date":"30 Aug 2025"}},{"key":"optimus-teleop-16","href":"/platforms/tesla-optimus","source":{"label":"Business Insider (via AOL)","url":"https://www.aol.com/tesla-shifted-optimus-training-strategy-092502010.html","date":"2025-08-25"},"cells":{"subject":"Optimus","value":"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.","date":"25 Aug 2025"}},{"key":"optimus-teleop-17","href":"/platforms/tesla-optimus","source":{"label":"Business Insider (via AOL)","url":"https://www.aol.com/tesla-shifted-optimus-training-strategy-092502010.html","date":"2025-08-25"},"cells":{"subject":"Optimus","value":"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.","date":"25 Aug 2025"}},{"key":"atlas-large-behavior-model-teleop-18","href":"/platforms/boston-dynamics","source":{"label":"Boston Dynamics","url":"https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/","date":"2025-08-20"},"cells":{"subject":"Atlas Large Behavior Model","value":"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.","date":"20 Aug 2025"}},{"key":"skild-brain-teleop-19","href":"/platforms/skild-ai","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain","date":"2025-07-29"},"cells":{"subject":"Skild Brain","value":"Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.","date":"29 Jul 2025"}},{"key":"gr-3-teleop-20","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/abs/2507.15493","date":"2025-07-21"},"cells":{"subject":"GR-3","value":"ByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.","date":"21 Jul 2025"}},{"key":"gr-3-teleop-21","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/html/2507.15493v1","date":"2025-07-21"},"cells":{"subject":"GR-3","value":"For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.","date":"21 Jul 2025"}},{"key":"gr-3-teleop-22","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/html/2507.15493v1","date":"2025-07-21"},"cells":{"subject":"GR-3","value":"ByteDance said human trajectories recorded with VR devices come in at about 450 per hour, against about 250 per hour for teleoperated robot trajectories.","date":"21 Jul 2025"}},{"key":"tri-lbm-teleop-23","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","value":"The real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.","date":"7 Jul 2025"}},{"key":"tri-lbm-teleop-24","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","value":"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.","date":"7 Jul 2025"}},{"key":"gr00t-n1-5-teleop-25","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_5/","date":"2025-06-11"},"cells":{"subject":"GR00T N1.5","value":"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.","date":"11 Jun 2025"}},{"key":"v-jepa-2-teleop-26","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2","value":"Meta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.","date":"11 Jun 2025"}},{"key":"redwood-teleop-27","href":"/platforms/1x","source":{"label":"1X Technologies","url":"https://www.1x.tech/discover/redwood-ai","date":"2025-06-10"},"cells":{"subject":"Redwood","value":"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.","date":"10 Jun 2025"}},{"key":"smolvla-teleop-28","source":{"label":"Hugging Face","url":"https://huggingface.co/blog/smolvla","date":"2025-06-03"},"cells":{"subject":"SmolVLA","value":"SmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all.","date":"3 Jun 2025"}},{"key":"gr00t-n1-5-teleop-29","source":{"label":"NVIDIA Newsroom","url":"https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai","date":"2025-05-18"},"cells":{"subject":"GR00T N1.5","value":"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.","date":"18 May 2025"}},{"key":"graspvla-teleop-30","source":{"label":"arXiv (Galbot, Peking University)","url":"https://arxiv.org/abs/2505.03233","date":"2025-05-06"},"cells":{"subject":"GraspVLA","value":"Galbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.","date":"6 May 2025"}},{"key":"pi0-5-teleop-31","href":"/platforms/physical-intelligence","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2504.16054v1","date":"2025-04-22"},"cells":{"subject":"π0.5","value":"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.","date":"22 Apr 2025"}},{"key":"gr00t-n1-teleop-32","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","value":"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.","date":"18 Mar 2025"}},{"key":"gr00t-n1-teleop-33","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","value":"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.","date":"18 Mar 2025"}},{"key":"gr00t-n1-teleop-34","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","value":"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.","date":"18 Mar 2025"}},{"key":"gemini-robotics-teleop-35","source":{"label":"arXiv (Google DeepMind)","url":"https://arxiv.org/html/2503.20020v1","date":"2025-03-25"},"cells":{"subject":"Gemini Robotics","value":"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.","date":"12 Mar 2025"}},{"key":"go-1-teleop-36","source":{"label":"arXiv (AgiBot)","url":"https://arxiv.org/html/2503.06669v4","date":"2025-08-04"},"cells":{"subject":"GO-1 (Genie Operator-1)","value":"AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.","date":"10 Mar 2025"}},{"key":"go-1-teleop-37","source":{"label":"Newsfile Corp (AgiBot release)","url":"https://www.newsfilecorp.com/release/243954","date":"2025-03-10"},"cells":{"subject":"GO-1 (Genie Operator-1)","value":"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.","date":"10 Mar 2025"}},{"key":"helix-teleop-38","href":"/platforms/figure","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix","date":"2025-02-20"},"cells":{"subject":"Helix","value":"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.","date":"20 Feb 2025"}},{"key":"gr-2-teleop-39","source":{"label":"arXiv (ByteDance Research)","url":"https://arxiv.org/abs/2410.06158","date":"2024-10-08"},"cells":{"subject":"GR-2","value":"ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.","date":"8 Oct 2024"}},{"key":"rfm-1-teleop-40","source":{"label":"MIT Technology Review","url":"https://www.technologyreview.com/2024/03/11/1089653/an-openai-spinoff-has-built-an-ai-model-that-helps-robots-learn-tasks-like-humans/","date":"2024-03-11"},"cells":{"subject":"RFM-1","value":"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.","date":"11 Mar 2024"}},{"key":"diffusion-policy-behavior-model-teleop-41","source":{"label":"Toyota USA Newsroom","url":"https://pressroom.toyota.com/toyota-research-institute-unveils-breakthrough-in-teaching-robots-new-behaviors/","date":"2023-09-19"},"cells":{"subject":"Diffusion Policy behavior model","value":"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.","date":"19 Sep 2023"}},{"key":"rt-2-teleop-42","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action/","date":"2023-07-28"},"cells":{"subject":"RT-2","value":"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.","date":"28 Jul 2023"}},{"key":"robocat-teleop-43","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/robocat-a-self-improving-robotic-agent/","date":"2023-06-20"},"cells":{"subject":"RoboCat","value":"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.","date":"20 Jun 2023"}}],"sample":"44 facts from 32 robot models.","notes":[]},{"id":"facts","title":"Every sourced fact","question":"What exactly has been published about each model?","columns":[{"key":"subject","label":"Model","unit":"text"},{"key":"kind","label":"Kind","unit":"text"},{"key":"fact","label":"Fact","unit":"text"},{"key":"date","label":"Date","unit":"text","secondary":true},{"key":"basis","label":"Basis","unit":"text","secondary":true}],"rows":[{"key":"1x-world-model-0","source":{"label":"1X Technologies","url":"https://1x.tech/discover/world-model-self-learning","date":"2026-01-12"},"cells":{"subject":"1X World Model (1XWM)","kind":"Human video","fact":"1X mid-trained its 1XWM world model on 900 hours of egocentric human video after pre-training on web-scale video.","date":"12 Jan 2026","basis":"Stated by the company or an official page"}},{"key":"1x-world-model-1","source":{"label":"1X Technologies","url":"https://1x.tech/discover/world-model-self-learning","date":"2026-01-12"},"cells":{"subject":"1X World Model (1XWM)","kind":"Hours of training data","fact":"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.","date":"12 Jan 2026","basis":"Stated by the company or an official page"}},{"key":"act-1-2","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/no-robot-data","date":"2025-11-19"},"cells":{"subject":"ACT-1","kind":"Teleoperation data","fact":"Sunday Robotics trained ACT-1 without a single trajectory of teleoperation data, learning instead from people wearing its Skill Capture Glove.","date":"19 Nov 2025","basis":"Stated by the company or an official page"}},{"key":"act-1-3","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/no-robot-data","date":"2025-11-19"},"cells":{"subject":"ACT-1","kind":"Data mix","fact":"Sunday converts data recorded with its Skill Capture Glove into equivalent robot training data with a 90% success rate.","date":"19 Nov 2025","basis":"Stated by the company or an official page"}},{"key":"act-2-4","source":{"label":"Sunday Robotics","url":"https://www.sunday.ai/blog/act-2-preview","date":"2026-07-17"},"cells":{"subject":"ACT-2","kind":"Data mix","fact":"Sunday pre-trains ACT-2 on a sensorized human dataset collected with its own data collection hardware.","date":"17 Jul 2026","basis":"Stated by the company or an official page"}},{"key":"atlas-large-behavior-model-5","source":{"label":"Boston Dynamics","url":"https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/","date":"2025-08-20"},"cells":{"subject":"Atlas Large Behavior Model","kind":"Teleoperation data","fact":"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.","date":"20 Aug 2025","basis":"Stated by the company or an official page"}},{"key":"diffusion-policy-behavior-model-6","source":{"label":"Toyota USA Newsroom","url":"https://pressroom.toyota.com/toyota-research-institute-unveils-breakthrough-in-teaching-robots-new-behaviors/","date":"2023-09-19"},"cells":{"subject":"Diffusion Policy behavior model","kind":"Teleoperation data","fact":"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.","date":"19 Sep 2023","basis":"Stated by the company or an official page"}},{"key":"dyna-2-7","source":{"label":"Dyna Robotics","url":"https://www.dyna.co/dyna-2","date":"2026-08-01"},"cells":{"subject":"Dyna-2","kind":"Human video","fact":"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.","date":"1 Aug 2026","basis":"Stated by the company or an official page"}},{"key":"dyna-2-8","source":{"label":"Dyna Robotics","url":"https://www.dyna.co/dyna-2","date":"2026-08-01"},"cells":{"subject":"Dyna-2","kind":"Teleoperation data","fact":"Dyna said 10 minutes of teleoperation data were enough to fine-tune Dyna-2 to open bottle caps with two five-fingered robot hands.","date":"1 Aug 2026","basis":"Stated by the company or an official page"}},{"key":"g0-9","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","kind":"Hours of training data","fact":"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.","date":"30 Aug 2025","basis":"Stated by the company or an official page"}},{"key":"g0-10","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","kind":"Teleoperation data","fact":"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.","date":"30 Aug 2025","basis":"Stated by the company or an official page"}},{"key":"g0-11","source":{"label":"arXiv (Galaxea)","url":"https://arxiv.org/html/2509.00576v1","date":"2025-08-30"},"cells":{"subject":"G0","kind":"Data mix","fact":"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.","date":"30 Aug 2025","basis":"Stated by the company or an official page"}},{"key":"gemini-robotics-12","source":{"label":"arXiv (Google DeepMind)","url":"https://arxiv.org/html/2503.20020v1","date":"2025-03-25"},"cells":{"subject":"Gemini Robotics","kind":"Teleoperation data","fact":"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.","date":"12 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gemini-robotics-1-5-13","source":{"label":"arXiv (Google DeepMind)","url":"https://arxiv.org/html/2510.03342v1","date":"2025-10-02"},"cells":{"subject":"Gemini Robotics 1.5","kind":"Robot types","fact":"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.","date":"25 Sep 2025","basis":"Stated by the company or an official page"}},{"key":"gemini-robotics-on-device-14","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/","date":"2025-06-24"},"cells":{"subject":"Gemini Robotics On-Device","kind":"Robot types","fact":"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.","date":"24 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"gen-0-15","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/nov-04-2025-GEN-0","date":"2025-11-04"},"cells":{"subject":"GEN-0","kind":"Hours of training data","fact":"Generalist said GEN-0 is pre-trained on over 270,000 hours of real-world manipulation data, growing by 10,000 hours a week.","date":"4 Nov 2025","basis":"Stated by the company or an official page"}},{"key":"gen-0-16","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/nov-04-2025-GEN-0","date":"2025-11-04"},"cells":{"subject":"GEN-0","kind":"Data mix","fact":"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.","date":"4 Nov 2025","basis":"Stated by the company or an official page"}},{"key":"gen-1-17","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","kind":"Hours of training data","fact":"Generalist trained GEN-1 from scratch on a dataset it put at over half a million hours of physical interaction data.","date":"2 Apr 2026","basis":"Stated by the company or an official page"}},{"key":"gen-1-18","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","kind":"Human video","fact":"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.","date":"2 Apr 2026","basis":"Stated by the company or an official page"}},{"key":"gen-1-19","source":{"label":"Generalist AI","url":"https://generalistai.com/blog/gen-1","date":"2026-04-02"},"cells":{"subject":"GEN-1","kind":"Hours of training data","fact":"Each GEN-1 task result Generalist showed was built with about one hour of robot data.","date":"2 Apr 2026","basis":"Stated by the company or an official page"}},{"key":"go-1-20","source":{"label":"arXiv (AgiBot)","url":"https://arxiv.org/html/2503.06669v4","date":"2025-08-04"},"cells":{"subject":"GO-1 (Genie Operator-1)","kind":"Episodes","fact":"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.","date":"10 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"go-1-21","source":{"label":"arXiv (AgiBot)","url":"https://arxiv.org/html/2503.06669v4","date":"2025-08-04"},"cells":{"subject":"GO-1 (Genie Operator-1)","kind":"Teleoperation data","fact":"AgiBot said the over 1 million trajectories in AgiBot World were gathered by skilled teleoperators, using VR headset and whole-body motion capture teleoperation.","date":"10 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"go-1-22","source":{"label":"Newsfile Corp (AgiBot release)","url":"https://www.newsfilecorp.com/release/243954","date":"2025-03-10"},"cells":{"subject":"GO-1 (Genie Operator-1)","kind":"Data mix","fact":"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.","date":"10 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gr-2-23","source":{"label":"arXiv (ByteDance Research)","url":"https://arxiv.org/abs/2410.06158","date":"2024-10-08"},"cells":{"subject":"GR-2","kind":"Human video","fact":"ByteDance pre-trained GR-2 on 38 million video clips, over 50 billion tokens, before fine-tuning on robot trajectories.","date":"8 Oct 2024","basis":"Stated by the company or an official page"}},{"key":"gr-3-24","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/abs/2507.15493","date":"2025-07-21"},"cells":{"subject":"GR-3","kind":"Data mix","fact":"ByteDance trains GR-3 with web-scale vision-language data, human trajectory data collected with VR devices, and robot trajectory data for imitation learning.","date":"21 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"gr-3-25","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/html/2507.15493v1","date":"2025-07-21"},"cells":{"subject":"GR-3","kind":"Teleoperation data","fact":"For one pick-and-place task, ByteDance collected 35,000 teleoperated robot trajectories covering 101 objects, 69 hours in total.","date":"21 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"gr-3-26","source":{"label":"arXiv (ByteDance Seed)","url":"https://arxiv.org/html/2507.15493v1","date":"2025-07-21"},"cells":{"subject":"GR-3","kind":"Data mix","fact":"ByteDance said human trajectories recorded with VR devices come in at about 450 per hour, against about 250 per hour for teleoperated robot trajectories.","date":"21 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-27","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","kind":"Teleoperation data","fact":"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.","date":"18 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-28","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","kind":"Data mix","fact":"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.","date":"18 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-29","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","kind":"Episodes","fact":"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.","date":"18 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-30","source":{"label":"arXiv (NVIDIA)","url":"https://arxiv.org/html/2503.14734v2","date":"2025-03-18"},"cells":{"subject":"GR00T N1","kind":"Human video","fact":"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.","date":"18 Mar 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-5-31","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_5/","date":"2025-06-11"},"cells":{"subject":"GR00T N1.5","kind":"Data mix","fact":"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.","date":"11 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-5-32","source":{"label":"NVIDIA Newsroom","url":"https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai","date":"2025-05-18"},"cells":{"subject":"GR00T N1.5","kind":"Data mix","fact":"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.","date":"18 May 2025","basis":"Stated by the company or an official page"}},{"key":"gr00t-n1-6-33","source":{"label":"NVIDIA Research (GEAR)","url":"https://research.nvidia.com/labs/gear/gr00t-n1_6/","date":"2025-12-15"},"cells":{"subject":"GR00T N1.6","kind":"Teleoperation data","fact":"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.","date":"15 Dec 2025","basis":"Stated by the company or an official page"}},{"key":"graspvla-34","source":{"label":"arXiv (Galbot, Peking University)","url":"https://arxiv.org/abs/2505.03233","date":"2025-05-06"},"cells":{"subject":"GraspVLA","kind":"Data mix","fact":"Galbot and its academic partners pre-trained GraspVLA entirely on synthetic action data: SynGrasp-1B, a billion-frame grasping dataset generated in simulation.","date":"6 May 2025","basis":"Stated by the company or an official page"}},{"key":"helix-35","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix","date":"2025-02-20"},"cells":{"subject":"Helix","kind":"Teleoperation data","fact":"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.","date":"20 Feb 2025","basis":"Stated by the company or an official page"}},{"key":"helix-project-go-big-36","source":{"label":"Figure AI","url":"https://www.figure.ai/news/project-go-big","date":"2025-09-18"},"cells":{"subject":"Helix (Project Go-Big)","kind":"Human video","fact":"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.","date":"18 Sep 2025","basis":"Stated by the company or an official page"}},{"key":"helix-02-37","source":{"label":"Figure AI","url":"https://www.figure.ai/news/helix-02","date":"2026-01-27"},"cells":{"subject":"Helix 02","kind":"Hours of training data","fact":"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.","date":"27 Jan 2026","basis":"Stated by the company or an official page"}},{"key":"tri-lbm-38","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","kind":"Hours of training data","fact":"TRI trained its Large Behavior Models on about 1,700 hours of robot demonstrations covering over 500 internally collected tasks plus public robot data.","date":"7 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"tri-lbm-39","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","kind":"Episodes","fact":"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.","date":"7 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"tri-lbm-40","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","kind":"Teleoperation data","fact":"The real-robot part of TRI's LBM data is 468 hours and 46,063 demonstrations over 362 tasks, collected across 9 hardware stations.","date":"7 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"tri-lbm-41","source":{"label":"arXiv (Toyota Research Institute)","url":"https://arxiv.org/html/2507.05331v1","date":"2025-07-07"},"cells":{"subject":"Large Behavior Model (LBM)","kind":"Data mix","fact":"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.","date":"7 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"octo-42","source":{"label":"arXiv","url":"https://arxiv.org/abs/2405.12213","date":"2024-05-20"},"cells":{"subject":"Octo","kind":"Episodes","fact":"Octo was trained on 800,000 robot trajectories from a mixture of 25 datasets in the Open X-Embodiment collection.","date":"20 May 2024","basis":"Stated by the company or an official page"}},{"key":"openvla-43","source":{"label":"arXiv","url":"https://arxiv.org/abs/2406.09246","date":"2024-06-13"},"cells":{"subject":"OpenVLA","kind":"Episodes","fact":"OpenVLA, a 7B-parameter open-source model, was trained on 970,000 real-world robot demonstrations curated from the Open X-Embodiment dataset.","date":"13 Jun 2024","basis":"Stated by the company or an official page"}},{"key":"optimus-44","source":{"label":"Business Insider (via AOL)","url":"https://www.aol.com/tesla-shifted-optimus-training-strategy-092502010.html","date":"2025-08-25"},"cells":{"subject":"Optimus","kind":"Human video","fact":"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.","date":"25 Aug 2025","basis":"Reported by an outlet"}},{"key":"optimus-45","source":{"label":"Business Insider (via AOL)","url":"https://www.aol.com/tesla-shifted-optimus-training-strategy-092502010.html","date":"2025-08-25"},"cells":{"subject":"Optimus","kind":"Teleoperation data","fact":"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.","date":"25 Aug 2025","basis":"Reported by an outlet"}},{"key":"rdt-1b-46","source":{"label":"RDT project (Tsinghua University)","url":"https://rdt-robotics.github.io/rdt-robotics/","date":"2026-10-18"},"cells":{"subject":"RDT-1B","kind":"Episodes","fact":"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.","date":"10 Oct 2024","basis":"Stated by the company or an official page"}},{"key":"redwood-47","source":{"label":"1X Technologies","url":"https://www.1x.tech/discover/redwood-ai","date":"2025-06-10"},"cells":{"subject":"Redwood","kind":"Teleoperation data","fact":"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.","date":"10 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"rfm-1-48","source":{"label":"MIT Technology Review","url":"https://www.technologyreview.com/2024/03/11/1089653/an-openai-spinoff-has-built-an-ai-model-that-helps-robots-learn-tasks-like-humans/","date":"2024-03-11"},"cells":{"subject":"RFM-1","kind":"Data mix","fact":"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.","date":"11 Mar 2024","basis":"Reported by an outlet"}},{"key":"rfm-1-49","source":{"label":"IEEE Spectrum","url":"https://spectrum.ieee.org/covariant-foundation-model","date":"2024-03-11"},"cells":{"subject":"RFM-1","kind":"Episodes","fact":"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.","date":"11 Mar 2024","basis":"Reported by an outlet"}},{"key":"robocat-50","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/robocat-a-self-improving-robotic-agent/","date":"2023-06-20"},"cells":{"subject":"RoboCat","kind":"Teleoperation data","fact":"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.","date":"20 Jun 2023","basis":"Stated by the company or an official page"}},{"key":"rt-2-51","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action/","date":"2023-07-28"},"cells":{"subject":"RT-2","kind":"Data mix","fact":"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.","date":"28 Jul 2023","basis":"Stated by the company or an official page"}},{"key":"rt-x-52","source":{"label":"arXiv (Open X-Embodiment Collaboration)","url":"https://arxiv.org/abs/2310.08864","date":"2023-10-13"},"cells":{"subject":"RT-X (Open X-Embodiment)","kind":"Robot types","fact":"The Open X-Embodiment dataset behind the RT-X models pools data from 22 different robots, collected with 21 institutions, covering 527 skills.","date":"3 Oct 2023","basis":"Stated by the company or an official page"}},{"key":"rt-x-53","source":{"label":"Google DeepMind","url":"https://deepmind.google/discover/blog/scaling-up-learning-across-many-different-robot-types/","date":"2023-10-03"},"cells":{"subject":"RT-X (Open X-Embodiment)","kind":"Episodes","fact":"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.","date":"3 Oct 2023","basis":"Stated by the company or an official page"}},{"key":"skild-brain-54","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/learning-by-watching","date":"2026-01-12"},"cells":{"subject":"Skild Brain","kind":"Human video","fact":"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.","date":"12 Jan 2026","basis":"Stated by the company or an official page"}},{"key":"skild-brain-55","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/omni-bodied","date":"2025-09-24"},"cells":{"subject":"Skild Brain","kind":"Robot types","fact":"Skild AI said it trained its omni-bodied brain to control 100,000 different simulated robots, over a millennium of simulated time.","date":"24 Sep 2025","basis":"Stated by the company or an official page"}},{"key":"skild-brain-56","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain","date":"2025-07-29"},"cells":{"subject":"Skild Brain","kind":"Data mix","fact":"Skild AI pre-trains its Skild Brain on large-scale simulation and internet video, then post-trains it on targeted real-world data.","date":"29 Jul 2025","basis":"Stated by the company or an official page"}},{"key":"skild-s1-57","source":{"label":"Skild AI","url":"https://www.skild.ai/blogs/s1","date":"2026-08-18"},"cells":{"subject":"Skild S1","kind":"Hours of training data","fact":"For S1, Skild AI ran a controlled study on subsets of its pre-training data pool ranging from 1,000 to 100,000 hours.","date":"18 Aug 2026","basis":"Stated by the company or an official page"}},{"key":"smolvla-58","source":{"label":"arXiv (Hugging Face)","url":"https://arxiv.org/html/2506.01844v1","date":"2025-06-02"},"cells":{"subject":"SmolVLA","kind":"Episodes","fact":"Hugging Face trained SmolVLA on fewer than 30,000 episodes, all from public, community-contributed LeRobot datasets.","date":"3 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"smolvla-59","source":{"label":"Hugging Face","url":"https://huggingface.co/blog/smolvla","date":"2025-06-03"},"cells":{"subject":"SmolVLA","kind":"Data mix","fact":"SmolVLA's pre-training set is 487 community datasets for the low-cost SO-100 arm, about 10 million frames in all.","date":"3 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"v-jepa-2-60","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2","kind":"Human video","fact":"Meta pre-trained V-JEPA 2 on a video and image dataset of over 1 million hours of internet video.","date":"11 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"v-jepa-2-ac-61","source":{"label":"arXiv (Meta FAIR)","url":"https://arxiv.org/abs/2506.09985","date":"2025-06-11"},"cells":{"subject":"V-JEPA 2-AC","kind":"Hours of training data","fact":"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.","date":"11 Jun 2025","basis":"Stated by the company or an official page"}},{"key":"pi-star-0-6-62","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pistar06","date":"2025-11-17"},"cells":{"subject":"π*0.6","kind":"Teleoperation data","fact":"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.","date":"17 Nov 2025","basis":"Stated by the company or an official page"}},{"key":"pi0-63","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/openpi","date":"2025-02-04"},"cells":{"subject":"π0","kind":"Release","fact":"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.","date":"4 Feb 2025","basis":"Stated by the company or an official page"}},{"key":"pi0-64","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2410.24164v1","date":"2024-10-31"},"cells":{"subject":"π0","kind":"Hours of training data","fact":"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.","date":"31 Oct 2024","basis":"Stated by the company or an official page"}},{"key":"pi0-65","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2410.24164v1","date":"2024-10-31"},"cells":{"subject":"π0","kind":"Robot types","fact":"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.","date":"31 Oct 2024","basis":"Stated by the company or an official page"}},{"key":"pi0-5-66","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2504.16054v1","date":"2025-04-22"},"cells":{"subject":"π0.5","kind":"Hours of training data","fact":"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.","date":"22 Apr 2025","basis":"Stated by the company or an official page"}},{"key":"pi0-5-67","source":{"label":"arXiv (Physical Intelligence)","url":"https://arxiv.org/html/2504.16054v1","date":"2025-04-22"},"cells":{"subject":"π0.5","kind":"Data mix","fact":"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.","date":"22 Apr 2025","basis":"Stated by the company or an official page"}},{"key":"pi0-5-human-to-robot-transfer-68","source":{"label":"Physical Intelligence","url":"https://www.pi.website/research/human_to_robot","date":"2025-12-16"},"cells":{"subject":"π0.5 (human-to-robot transfer)","kind":"Human video","fact":"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.","date":"16 Dec 2025","basis":"Stated by the company or an official page"}},{"key":"pi0-7-69","source":{"label":"Physical Intelligence","url":"https://www.pi.website/blog/pi07","date":"2026-04-16"},"cells":{"subject":"π0.7","kind":"Data mix","fact":"π0.7 is trained on data from many different robots, human data, and autonomous episodes collected by running various policies.","date":"16 Apr 2026","basis":"Stated by the company or an official page"}}],"sample":"All 70 facts. \"Reported\" means an outlet reports it from sources the company has not confirmed.","notes":["Amounts stay in the currency and form their source gives; nothing is converted or totalled."]}]}