TechniaHQRobot logoTechniaHQRobot
Training data

Robot Learning Data Pipeline

Fourteen records retain their original units and task scope. Hours, trajectories, robots and simulation environments are not plotted on a shared scale.

Source observations through 2026-07-31. Analysis revised separately.Open all humanoid data charts

Data summary

What is included

Fourteen records retain their original units and task scope. Hours, trajectories, robots and simulation environments are not plotted on a shared scale.

Visible rows

14

Dataset scope

TechniaHQRobot source-linked dataset

Interactive infographic

Robot Learning Data Pipeline

Fourteen records retain their original units and task scope. Hours, trajectories, robots and simulation environments are not plotted on a shared scale.

  1. Figure 02

    Recorded quantity
    500 hours
    Collection scope
    Multi-robot, multi-operator teleoperation data
    Robot or dataset
    Figure 02

    Published for the original Helix system.

    Read source ↗

  2. Figure 02

    Recorded quantity
    10 hours
    Collection scope
    Moving-conveyor package manipulation
    Robot or dataset
    Figure 02

    One of four controlled training-volume conditions in the scaling study.

    Read source ↗

  3. Figure 02

    Recorded quantity
    20 hours
    Collection scope
    Moving-conveyor package manipulation
    Robot or dataset
    Figure 02

    One of four controlled training-volume conditions in the scaling study.

    Read source ↗

  4. Figure 02

    Recorded quantity
    40 hours
    Collection scope
    Moving-conveyor package manipulation
    Robot or dataset
    Figure 02

    One of four controlled training-volume conditions in the scaling study.

    Read source ↗

  5. Figure 02

    Recorded quantity
    60 hours
    Collection scope
    Moving-conveyor package manipulation
    Robot or dataset
    Figure 02

    One of four controlled training-volume conditions in the scaling study.

    Read source ↗

  6. Figure 03

    Recorded quantity
    More than 1,000 hours
    Collection scope
    Human motion data used for full-body control
    Robot or dataset
    Figure 03

    The source states more than 1,000 hours.

    Read source ↗

  7. Figure 03

    Recorded quantity
    200,000 parallel environments
    Collection scope
    Simulation environments used for locomotion/control training
    Robot or dataset
    Figure 03

    Environment count is not a trajectory or hour count.

    Read source ↗

  8. Five robot types

    Recorded quantity
    1,000,000 trajectories
    Collection scope
    More than 1,000 tasks collected with more than 100 robots
    Robot or dataset
    Five robot types

    Trajectories are not equivalent to hours or validated task successes.

    Read source ↗

  9. Five robot types

    Recorded quantity
    More than 100 robots
    Collection scope
    Collection fleet lower bound
    Robot or dataset
    Five robot types

    The source states more than 100 robots.

    Read source ↗

  10. Five robot types

    Recorded quantity
    More than 1,000 tasks
    Collection scope
    Task coverage lower bound
    Robot or dataset
    Five robot types

    The source states more than 1,000 tasks.

    Read source ↗

  11. Fourier humanoid platforms

    Recorded quantity
    More than 30,000 trajectories
    Collection scope
    Bimanual teleoperation trajectories
    Robot or dataset
    Fourier humanoid platforms

    The source states more than 30,000 trajectories.

    Read source ↗

  12. NEO/EVE learning stack

    Recorded quantity
    900 hours
    Collection scope
    Egocentric human demonstrations
    Robot or dataset
    NEO/EVE learning stack

    Different modality from robot teleoperation; do not add without a unit-aware breakdown.

    Read source ↗

  13. NEO/EVE learning stack

    Recorded quantity
    70 hours
    Collection scope
    Robot fine-tuning data
    Robot or dataset
    NEO/EVE learning stack

    Reported separately from human video and unfiltered robot data.

    Read source ↗

  14. NEO/EVE learning stack

    Recorded quantity
    400 hours
    Collection scope
    Unfiltered robot data
    Robot or dataset
    NEO/EVE learning stack

    Reported separately from fine-tuning data.

    Read source ↗

Accessible data table

The table contains the same source-linked records used by the chart. It is the reference when labels overlap or a visual scale compresses values.

Source-linked data used in the infographic
Pipeline stageOrderDocumented volumeUnitStatusRobot or datasetCollection scopeRecord notesEvidenceSource
Figure Helix multi-robot teleoperation dataset1500 hourshoursofficially-documentedFigure 02Multi-robot, multi-operator teleoperation dataPublished for the original Helix system.Official announcementFigure Helix official technical articlePublished 2025-02-20Checked 2026-07-31
Figure Helix logistics training subset210 hourshoursofficially-documentedFigure 02Moving-conveyor package manipulationOne of four controlled training-volume conditions in the scaling study.Official announcementFigure official Helix logistics scaling studyPublished 2025-06-07Checked 2026-07-31
Figure Helix logistics training subset220 hourshoursofficially-documentedFigure 02Moving-conveyor package manipulationOne of four controlled training-volume conditions in the scaling study.Official announcementFigure official Helix logistics scaling studyPublished 2025-06-07Checked 2026-07-31
Figure Helix logistics training subset240 hourshoursofficially-documentedFigure 02Moving-conveyor package manipulationOne of four controlled training-volume conditions in the scaling study.Official announcementFigure official Helix logistics scaling studyPublished 2025-06-07Checked 2026-07-31
Figure Helix logistics training subset260 hourshoursofficially-documentedFigure 02Moving-conveyor package manipulationOne of four controlled training-volume conditions in the scaling study.Official announcementFigure official Helix logistics scaling studyPublished 2025-06-07Checked 2026-07-31
Figure Helix 02 human-motion retargeting3More than 1,000 hourshoursofficially-documentedFigure 03Human motion data used for full-body controlThe source states more than 1,000 hours.Official announcementFigure Helix 02 official technical articlePublished 2026-01-27Checked 2026-07-31
Figure Helix 02 parallel simulation4200,000 parallel environmentsparallel environmentsofficially-documentedFigure 03Simulation environments used for locomotion/control trainingEnvironment count is not a trajectory or hour count.Official announcementFigure Helix 02 official technical articlePublished 2026-01-27Checked 2026-07-31
AgiBot World embodied dataset51,000,000 trajectoriestrajectoriesofficially-documentedFive robot typesMore than 1,000 tasks collected with more than 100 robotsTrajectories are not equivalent to hours or validated task successes.Official announcementAgiBot official research pageChecked 2026-07-31
AgiBot World collection fleet5More than 100 robotsrobotsofficially-documentedFive robot typesCollection fleet lower boundThe source states more than 100 robots.Official announcementAgiBot official research pageChecked 2026-07-31
AgiBot World task coverage5More than 1,000 taskstasksofficially-documentedFive robot typesTask coverage lower boundThe source states more than 1,000 tasks.Official announcementAgiBot official research pageChecked 2026-07-31
Fourier ActionNet bimanual teleoperation dataset6More than 30,000 trajectoriestrajectoriesofficially-documentedFourier humanoid platformsBimanual teleoperation trajectoriesThe source states more than 30,000 trajectories.Official announcementFourier official teleoperation and ActionNet documentationChecked 2026-07-31
1X egocentric human video7900 hourshoursofficially-documentedNEO/EVE learning stackEgocentric human demonstrationsDifferent modality from robot teleoperation; do not add without a unit-aware breakdown.Official announcement1X official world-model training articleChecked 2026-07-31
1X robot fine-tuning data870 hourshoursofficially-documentedNEO/EVE learning stackRobot fine-tuning dataReported separately from human video and unfiltered robot data.Official announcement1X official world-model training articleChecked 2026-07-31
1X unfiltered robot data8400 hourshoursofficially-documentedNEO/EVE learning stackUnfiltered robot dataReported separately from fine-tuning data.Official announcement1X official world-model training articleChecked 2026-07-31

Methodology

How the dataset is built

Hours, trajectories, robots, tasks and simulation environments are different units and are never summed into one volume. Each row preserves the collection modality and scope stated by the source.

Editorial analysis

The records describe different training resources

The table combines teleoperation, human-motion data, simulated environments and embodied datasets. An hour measures duration. A trajectory measures an episode. A fleet count describes collection hardware. None can be converted to another without additional information.

Subsets can overlap larger collections

The four Figure logistics training subsets belong to one scaling study. They should not be added as independent datasets without establishing that their episodes are disjoint. Likewise, a robot count and task count can describe the same collection rather than extra training volume.

Volume leaves quality unresolved

For a training decision, inspect task coverage, recording quality, action labels and compatibility with the target robot. A larger number of trajectories can still contain repetitive or unsuitable examples. This page organizes the reported resources without assigning a universal data-quality score.

Frequently asked questions

Why is there no largest-dataset leaderboard?

The records use incompatible units and may overlap. A single numerical ranking would conceal those differences.

Sources

Related humanoid infographics

Humanoid Robot Companies by CountryHumanoid Robot Price DistributionHumanoid Robot Price vs HeightRobot Mass vs Payload

Continue with the source-linked humanoid catalog

Open individual robot records for manufacturer, origin, dimensions, availability, public price status and official product links.

Browse humanoid robotsRead the main guide