Skip to content

Research · Compiled from public sources

Open Robot Training Datasets: Size, Collection Method and Licence

32 open datasets of robot, first-person video, driving and drone training data were released from 2023 on; each is listed with who released it, its stated size, how its data was collected and its licence, linked to its source.

Covers
2 Jan 2023 to 18 Oct 2026
Dated by when each event occurred
Calculated
18 Oct 2026, 20:07 UTC
Sources last checked 18 Oct 2026
Where it comes from
Each dataset's paper, project page or dataset card, and reporting where the release gives no date or worker count, each linked with a verbatim quote kept in the dataset.
Method
Revision 1. Terms defined
CC BY 4.0, citing this report
Every figure as CSVJSON

What the data shows

  1. 1

    32 open datasets since 2023: 16 robot manipulation, 10 first-person video, 4 driving and 2 drones.

    Figure 2
  2. 2

    The largest by stated hours is Egocentric-100K (Build AI, 19 Dec 2025): 100,405 hours.

    Figure 3
  3. 3

    12 of the 32 were collected by people teleoperating robots, the work this site's teleoperation listings pay for.

    Figure 1
  4. 4

    The newest release on record: EgoSuite-Open100K (26 Aug 2026). Lightwheel and Hugging Face opened the first 10,000 hours of a planned 100,000-hour egocentric human activity dataset recorded with head-mounted cameras (plus a wrist camera in the EgoPro subset), with hand and body pose annotations.

    Figure 1

Figure 1

Open training datasets released since 2023

Which public datasets teach robots, embodied AI, self-driving systems and drones, who released them, how big are they and how was the data collected?

DatasetReleased byKindCollected by
EgoSuite-Open100KLightwheel (Hugging Face blog)Lightwheel with Hugging FaceFirst-person videoWearable camera
BEHAVIOR Challenge 2026 demonstrationsStanford Vision and Learning LabStanford Vision and Learning Lab (BEHAVIOR)Robot manipulationTeleoperation
HIW-500 (Humanoids-in-the-Wild 500)BitRobot (Hugging Face)BitRobot Network with Hugging Face and Unitree RoboticsRobot manipulationTeleoperation
ABC-130KarXiv (UC Berkeley, MIT, Amazon FAR, XDOF, CMU)XDOF with UC Berkeley, MIT, Amazon FAR and Carnegie Mellon UniversityRobot manipulationTeleoperation
MolmoAct 2-Bimanual YAMAi2Ai2 (Allen Institute for AI)Robot manipulationNot stated or other
Egocentric-100KBuild AI (Hugging Face)Build AIFirst-person videoWearable camera
RoboCOINarXiv (Beijing Academy of Artificial Intelligence)Beijing Academy of Artificial Intelligence (BAAI) and partner institutionsRobot manipulationTeleoperation
Egocentric-10KBuild AI (Hugging Face)Build AIFirst-person videoWearable camera
Waymo Open Dataset for End-to-End Driving (WOD-E2E)arXiv (Waymo)WaymoDrivingNot stated or other
Humanoid EverydayarXiv (USC, Toyota Research Institute)University of Southern California and Toyota Research InstituteRobot manipulationTeleoperation
Galaxea Open-World DatasetarXiv (Galaxea AI)Galaxea AIRobot manipulationTeleoperation
MolmoAct DatasetAi2Ai2 (Allen Institute for AI)Robot manipulationNot stated or other
LeRobot community datasets (SmolVLA pretraining set)Hugging FaceHugging Face (LeRobot) and community contributorsRobot manipulationMixed
UAV-FlowarXiv (Beihang University, NUS)Beihang University and National University of SingaporeDronesTeleoperation
EgoDexarXiv (Apple)AppleFirst-person videoWearable camera
Fourier ActionNetFourier (ActionNet project page)FourierRobot manipulationTeleoperation
HD-EPICarXiv (University of Bristol and partners)University of Bristol, Leiden University, Singapore Management University and University of Bath (EPIC-KITCHENS team)First-person videoWearable camera
AgiBot WorldarXiv (AgiBot World Colosseo team)AgiBotRobot manipulationTeleoperation
RoboMINDarXiv (Beijing Innovation Center of Humanoid Robotics, Peking University)Beijing Innovation Center of Humanoid Robotics (X-Humanoid), Peking University and Beijing Academy of Artificial IntelligenceRobot manipulationTeleoperation
MAN TruckScenesarXiv (MAN Truck & Bus, TUM)MAN Truck & Bus with the Technical University of MunichDrivingNot stated or other
CityNavarXivTokyo Institute of Technology, The University of Tokyo, Kyoto University and partnersDronesNot stated or other
NymeriaarXiv (Meta Reality Labs Research)Meta (Reality Labs Research, Project Aria)First-person videoWearable camera
HOT3DarXiv (Meta)Meta (Reality Labs)First-person videoWearable camera
DROID (Distributed Robot Interaction Dataset)arXiv (DROID collaboration)DROID collaboration (13 institutions)Robot manipulationTeleoperation
Aria Everyday Activities (AEA)arXiv (Meta)Meta (Project Aria)First-person videoWearable camera
Ego-Exo4DarXiv (Meta FAIR and university partners)Meta (FAIR and Project Aria) with 14 university partnersFirst-person videoMixed
Open X-EmbodimentarXiv (Open X-Embodiment Collaboration)Google DeepMind and the Open X-Embodiment Collaboration (21 institutions)Robot manipulationMixed
BridgeData V2arXiv (UC Berkeley, Stanford)UC Berkeley (RAIL) with Stanford and collaboratorsRobot manipulationMixed
RH20TarXiv (Shanghai Jiao Tong University)Shanghai Jiao Tong UniversityRobot manipulationTeleoperation
Zenseact Open Dataset (ZOD)arXiv (Zenseact)ZenseactDrivingNot stated or other
Ego4D (version 2)Ego4DMeta (FAIR) and the Ego4D university consortiumFirst-person videoWearable camera
Argoverse 2arXiv (Argo AI)Argo AIDrivingNot stated or other

Base: 32 datasets released from 2023 to 18 Oct 2026, each linked to its paper or release page (reviewed 18 Oct 2026).

Hours and episodes are as each release states them; a blank means the release gives no figure. Episodes, trajectories and demonstrations are counted as one.

Credit: compiled by RobotTraining.jobs Research, Open Robot Training Datasets: Size, Collection Method and Licence, checked 18 Oct 2026.

CSV

Figure 2

Datasets by what they teach

How many open datasets are there for each kind of robot learning, and how was their data gathered?

KindDatasetsTeleoperatedWearable camera
Robot manipulation
16
110
First-person video
10
09
Driving
4
00
Drones
2
10

Base: 32 datasets.

Collected by: how the release says the data was gathered. Teleoperation means people steering robots remotely or in VR; wearable camera means people filming their own hands and tasks.

Credit: compiled by RobotTraining.jobs Research, Open Robot Training Datasets: Size, Collection Method and Licence, checked 18 Oct 2026.

CSV

Figure 3

The largest datasets by hours of recorded data

Which open datasets hold the most hours of human-made robot training data?

DatasetHours
Egocentric-100KBuild AI (Hugging Face)
100,405
EgoSuite-Open100KLightwheel (Hugging Face blog)
10,000
Egocentric-10KBuild AI (Hugging Face)
10,000
Ego4D (version 2)Ego4D
3,670
ABC-130KarXiv (UC Berkeley, MIT, Amazon FAR, XDOF, CMU)
3,500
AgiBot WorldarXiv (AgiBot World Colosseo team)
2,976
BEHAVIOR Challenge 2026 demonstrationsStanford Vision and Learning Lab
1,950
Ego-Exo4DarXiv (Meta FAIR and university partners)
1,286
EgoDexarXiv (Apple)
829
MolmoAct 2-Bimanual YAMAi2
720
HIW-500 (Humanoids-in-the-Wild 500)BitRobot (Hugging Face)
500
Galaxea Open-World DatasetarXiv (Galaxea AI)
500
DROID (Distributed Robot Interaction Dataset)arXiv (DROID collaboration)
350
RoboMINDarXiv (Beijing Innovation Center of Humanoid Robotics, Peking University)
306
NymeriaarXiv (Meta Reality Labs Research)
300
Fourier ActionNetFourier (ActionNet project page)
140
HD-EPICarXiv (University of Bristol and partners)
41
MolmoAct DatasetAi2
22
Waymo Open Dataset for End-to-End Driving (WOD-E2E)arXiv (Waymo)
12
Aria Everyday Activities (AEA)arXiv (Meta)
8

Base: 20 of 32 datasets state their hours.

Hours are not totalled across datasets: releases count them differently (raw video, cleaned episodes, approved data).

Credit: compiled by RobotTraining.jobs Research, Open Robot Training Datasets: Size, Collection Method and Licence, checked 18 Oct 2026.

CSV

Figure 4

Size, contributors and licence of each dataset

What does each dataset contain, who recorded it and under what licence can it be used?

DatasetSize as stated
EgoSuite-Open100KLightwheel (Hugging Face blog)100,000 hours planned; first 10,000 hours available; 15,000+ tasks and 15,000+ real-world collection scenes
BEHAVIOR Challenge 2026 demonstrationsStanford Vision and Learning Lab20,000 human teleoperation demos, 1,950 hours in total (in simulation)
HIW-500 (Humanoids-in-the-Wild 500)BitRobot (Hugging Face)500+ hours, 23K+ episodes and 10+ TB of raw data from 12 real homes
ABC-130KarXiv (UC Berkeley, MIT, Amazon FAR, XDOF, CMU)3,500 hours over 130K episodes across 195 tasks (project page: 134,806 episodes, 3,553 hours), plus 400 hours of sim-teleop data
MolmoAct 2-Bimanual YAMAi2over 720 hours of training demonstrations (the blog also calls it a 700-hour collection); 32,246 episodes on the Hugging Face card
Egocentric-100KBuild AI (Hugging Face)100,405 hours, 10.8 billion frames, 2,010,759 clips, 24.79 TB
RoboCOINarXiv (Beijing Academy of Artificial Intelligence)over 180,000 demonstrations from 15 robotic platforms, 421 bimanual tasks, 432 objects, 16 environments
Egocentric-10KBuild AI (Hugging Face)10,000 hours, 1.08 billion frames, 192,900 clips, 16.4 TB
Waymo Open Dataset for End-to-End Driving (WOD-E2E)arXiv (Waymo)4,021 driving segments (approximately 12 hours) of 20 seconds each, curated for long-tail scenarios
Humanoid EverydayarXiv (USC, Toyota Research Institute)10.3k trajectories and over 3 million frames across 260 tasks in 7 categories
Galaxea Open-World DatasetarXiv (Galaxea AI)500 hours, 100K demonstration trajectories, more than 150 tasks across 50 real-world scenes
MolmoAct DatasetAi2~10,000 distinct robot episodes (the MolmoAct 2 announcement describes 22 hours, about 10.6K trajectories)
LeRobot community datasets (SmolVLA pretraining set)Hugging Face487 curated community datasets focused on the SO100 robot arm, standardized at 30 FPS, around 10 million frames (the SmolVLA paper counts 481 datasets)
UAV-FlowarXiv (Beihang University, NUS)real-world dataset with 30k flight trajectories, plus a simulation dataset
EgoDexarXiv (Apple)829 hours of 1080p, 30 Hz egocentric video with 338,000 episodes across 194 tasks (90 million frames)
Fourier ActionNetFourier (ActionNet project page)over 30K teleoperated trajectories or around 140 hours of interaction data
HD-EPICarXiv (University of Bristol and partners)41 hours of video in 9 kitchens, 69 recipes, 59K fine-grained actions, 51K audio events, 20K object movements
AgiBot WorldarXiv (AgiBot World Colosseo team)1,001,552 trajectories totalling 2,976.4 hours, 217 tasks, 87 skills and 106 scenes (latest version per the paper); the GitHub Beta release lists 1,003,672 trajectories (~43.8 TB)
RoboMINDarXiv (Beijing Innovation Center of Humanoid Robotics, Peking University)107k demonstration trajectories (305.5 hours) across 479 tasks and 96 object classes on 4 robot embodiments, plus 5k real-world failure demonstrations
MAN TruckScenesarXiv (MAN Truck & Bus, TUM)more than 740 scenes of 20 s each, with 4 cameras, 6 lidar and 6 radar sensors
CityNavarXiv32,637 human demonstration trajectories paired with language descriptions, over 4.65 km2 of Cambridge and Birmingham
NymeriaarXiv (Meta Reality Labs Research)300 hours of daily activities from 264 participants across 50 locations, over 399 km travelled
HOT3DarXiv (Meta)over 833 minutes (more than 3.7M images) of multi-view image streams, 19 subjects, 33 rigid objects
DROID (Distributed Robot Interaction Dataset)arXiv (DROID collaboration)76k demonstration trajectories or 350 hours of interaction data, across 564 scenes and 84 tasks (the project page says 86 tasks)
Aria Everyday Activities (AEA)arXiv (Meta)143 daily activity sequences, over 7.5 accumulated hours, recorded in 5 locations
Ego-Exo4DarXiv (Meta FAIR and university partners)1,286 hours of video combined (1,286.3 ego and exo hours) from 740 participants in 13 cities and 123 scene contexts
Open X-EmbodimentarXiv (Open X-Embodiment Collaboration)1M+ real robot trajectories spanning 22 robot embodiments, pooled from 60 existing robot datasets; 527 skills (160,266 tasks)
BridgeData V2arXiv (UC Berkeley, Stanford)60,096 trajectories across 24 environments on a WidowX 250 arm: 50,365 teleoperated demonstrations and 9,731 rollouts from a scripted pick-and-place policy
RH20TarXiv (Shanghai Jiao Tong University)over 110,000 contact-rich robot manipulation sequences, each paired with a human demonstration video; tasks drawn from RLBench (48), MetaWorld (29) and 70 self-proposed tasks
Zenseact Open Dataset (ZOD)arXiv (Zenseact)100k curated camera frames, 1,473 sequences of 20 seconds and 29 drives of a few minutes, collected over two years in 14 European countries
Ego4D (version 2)Ego4D3,670 hours of daily-life video from 923 participants in 74 locations across 9 countries
Argoverse 2arXiv (Argo AI)Sensor Dataset: 1,000 annotated sequences; Lidar Dataset: 20,000 sequences; Motion Forecasting Dataset: 250,000 scenarios; six US cities

Base: All 32 datasets.

A licence is summarised as the release page gives it; read the licence itself before using a dataset.

Credit: compiled by RobotTraining.jobs Research, Open Robot Training Datasets: Size, Collection Method and Licence, checked 18 Oct 2026.

CSV

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

Each dataset's paper, project page or dataset card, and reporting where the release gives no date or worker count, 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.

Episode
One recorded attempt at a task (also called a trajectory or demonstration).
Egocentric
First-person video filmed from a camera the person wears, usually on the head.

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.
  • Releases count hours and episodes differently, so sizes are compared only as each release states them.
  • Datasets that are mostly synthetic, or released before 2023, are left out.

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.

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). Open Robot Training Datasets: Size, Collection Method and Licence (live data, 2 Jan 2023 to 18 Oct 2026). https://robottraining.jobs/research/open-robot-training-datasets (accessed 18 Oct 2026)

BibTeX

@misc{robottrainingjobs_open_robot_training_datasets_live,
  author = {{RobotTraining.jobs}},
  title = {Open Robot Training Datasets: Size, Collection Method and Licence (live data, 2 Jan 2023 to 18 Oct 2026)},
  year = {2026},
  howpublished = {\url{https://robottraining.jobs/research/open-robot-training-datasets}},
  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

Browse all 80 reports