Skip to content
humanoidrobots.training

datasets.humanoidrobots.training · checked 2026-10-11

Open humanoid datasets

Humanoids only: real whole-body robots under teleoperation, large humanoid corpora, human motion capture retargeted to humanoid joints for whole-body tracking in sim, and humanoid sim benchmarks. Every Hugging Face id answered the Hub API, every site answered 200, and every license is copied from the source, including the non-commercial ones. Episode, frame and fps counts are read from each dataset's own meta/info.json, or are the count the source states (noted); blank means the source does not state it.

Datasets
22
LeRobot format
10
Episodes (stated)
1,130,291
Non-commercial
6

Real humanoid teleoperation · 12

Most load with LeRobot

Real humanoid teleoperation datasets
DatasetEmbodimentLicenseFormatEpisodesFramesfpsHF dl 30dLinks
Unitree G1 whole-body + Inspire hands: clothes into washing machineunitreerobotics/G1_WBT_Inspire_Put_Clothes_into_Washing_MachineLoad clothes into a washing machineUnitree G1 with Inspire hands (robot_type unitree_g1_inspire)Unitree G1 →apache-2.0LeRobot v3.0 · parquet + mp4300486,315304,293Hugging Face ↗
Unitree G1 whole-body + BrainCo hands: plates into dishwasherunitreerobotics/G1_WBT_Brainco_Collect_Plates_Into_DishwasherCollect plates and load a dishwasherUnitree G1 with BrainCo hands (robot_type unitree_g1_brainco)Unitree G1 →apache-2.0LeRobot v3.0 · parquet + mp4300426,897305,135Hugging Face ↗
Unitree G1 + Dex3: toasted breadunitreerobotics/G1_Dex3_ToastedBread_DatasetBimanual toast handlingUnitree G1 with Dex3 hands (robot_type Unitree_G1)Unitree G1 →apache-2.0LeRobot v3.0 · parquet + mp4418352,022302,599Hugging Face ↗
Unitree G1 + Dex1: stack blocksunitreerobotics/G1_Dex1_Stack_BlockStack blocksUnitree G1 with Dex1 grippers (robot_type unitree_g1)Unitree G1 →apache-2.0LeRobot v3.0 · parquet + mp4690284,191303,387Hugging Face ↗
Unitree G1 + Dex1: fold towelunitreerobotics/G1_Dex1_Fold_TowelFold a towelUnitree G1 with Dex1 grippers (robot_type G1_ALL)Unitree G1 →apache-2.0LeRobot v2.1 · parquet + mp4200310,671301,663Hugging Face ↗
Unitree G1 + BrainCo hands: pick appleunitreerobotics/G1_Brainco_PickApple_DatasetPick an appleUnitree G1 with BrainCo dexterous hands (robot_type Unitree_G1_Brainco)Unitree G1 →apache-2.0LeRobot v3.0 · parquet + mp4200153,668301,741Hugging Face ↗
Unitree H1: fold clotheslerobot/unitreeh1_fold_clothesFold clothesUnitree H1 (info.json robot_type is unknown; the id names the robot)Unitree H1 →apache-2.0LeRobot v3.0 · parquet + mp43819,00050918Hugging Face ↗
Unitree H1: warehouselerobot/unitreeh1_warehouseWarehouse pick and placeUnitree H1 (info.json robot_type is unknown; the id names the robot)Unitree H1 →apache-2.0LeRobot v3.0 · parquet + mp42411,27550647Hugging Face ↗
GR00T Teleop G1: fruit pick-and-placenvidia/PhysicalAI-Robotics-GR00T-Teleop-G1Pick the named fruit (apple, pear, grapes, starfruit) and place itUnitree G1 with tri-finger hands, upper-body teleoperation (43-dim full-body + hand joint state and action)Unitree G1 →cc-by-4.0MP4 + HDF51,000·204,785Hugging Face ↗repo ↗
GR00T Teleop GR1 (DreamDojo release)nvidia/PhysicalAI-Robotics-GR00T-Teleop-GR1Teleoperated GR1 manipulation, released with the DreamDojo world modelFourier GR1 (GR1_robot/ subset, robot_type GR1T1)Fourier GR-1 →cc-by-nc-4.0LeRobot v2.0 · parquet + mp422,2096,362,2932059,832Hugging Face ↗dreamdojo-world.github.io ↗
Humanoid EverydayUSC-PSI-Lab/humanoid-everyday260+ tasks: dexterous manipulation, human-humanoid interaction, loco-manipulationUnitree G1 and H1 (info.json robot_type mixed; the lab also publishes separate -G1 and -H1 repos)Unitree G1 →Unitree H1 →apache-2.0LeRobot v2.1 · parquet + mp48,9493,436,1713055,096Hugging Face ↗humanoideveryday.github.io ↗repo ↗
Fourier ActionNetFourierIntelligence/ActionNetBimanual dexterous tabletop manipulationFourier GR1-T1, GR1-T2 and GR2 with Fourier dexterous hands, per the cardFourier GR-1 →Fourier GR-2 →cc-by-nc-sa-4.0gatedHDF5 + mp4 (converter to LeRobot v2 in fourier-lerobot)···9,328Hugging Face ↗action-net.org ↗repo ↗

Large humanoid corpora · 2

Large humanoid corpora datasets
DatasetEmbodimentLicenseFormatEpisodesFramesfpsHF dl 30dLinks
AgiBot World Betaagibot-world/AgiBotWorld-BetaReal-world manipulation across 100+ scenes in 5 domains, 200+ task typesAgiBot G1 (wheeled dual-arm humanoid), per the repoCC BY-NC-SA 4.0gatedStated in the gated card's agreement and the repo README; the card has no license field. Non-commercial.Per the dataset card (gated)1,003,672··189,911Hugging Face ↗agibot-world.com ↗repo ↗
AgiBot World Alphaagibot-world/AgiBotWorld-AlphaCurated subset of AgiBot World Beta, 36 tasksAgiBot G1 (wheeled dual-arm humanoid), per the repoCC BY-NC-SA 4.0gatedStated in the gated card's agreement and the repo README; the card has no license field. Non-commercial.Per the dataset card (gated)92,214··49,510Hugging Face ↗agibot-world.com ↗repo ↗

Motion capture for whole-body tracking · 6

Retarget, then track in sim, e.g. ProtoMotions

Motion capture for whole-body tracking datasets
DatasetEmbodimentLicenseFormatEpisodesFramesfpsHF dl 30dLinks
AMASSThe standard human motion source for humanoid motion-tracking and imitation RLHuman motion capture, 15 optical mocap datasets unified in the SMPL body modelNon-commercial scientific research only (MPI AMASS license)Free sign-in on the site is required to download. The license forbids commercial use, redistribution, and training networks for commercial use (amass.is.tue.mpg.de/license.html).SMPL body-model representation, per the site····amass.is.tue.mpg.de ↗
LAFAN1 (Ubisoft La Forge Animation)Walking, dancing and other themed mocap takes; the source of the retargeted LAFAN1 sets belowHuman motion capture, 5 subjects (Ubisoft, shot May 2017)CC BY-NC-ND 4.0Stated in the repo README (license.txt); the accompanying code is MIT. GitHub reports NOASSERTION.BVH77496,67230·repo ↗
LAFAN1 retargeted to Unitree G1 / H1 / H1-2lvhaidong/LAFAN1_Retargeting_DatasetLAFAN1 clips retargeted by interaction-mesh optimisation and IKUnitree G1, H1 and H1-2 joint trajectoriesUnitree G1 →Unitree H1 →Unitree H1-2 →not statedThe card has no license field. The repo's LICENSE file is BSD 3-Clause (copyright Unitree Robotics); the card notes the source LAFAN1 data is CC BY-NC-ND 4.0.CSV joint configurations (root pose + joints), 30 fps for G1···26,988Hugging Face ↗
Retargeted AMASS for robotics (Unitree G1)fleaven/Retargeted_AMASS_for_roboticsAMASS motions retargeted to G1 within the official joint limitsUnitree G1, 29 DoF (g1_29dof_rev_1_0)Unitree G1 →cc-by-4.0Card license; the card also says the data keeps AMASS's original terms, which are non-commercial.Per-clip arrays [-1, 36]: root position, root quaternion, 29 joint positions···3,779Hugging Face ↗
AMASS retargeted for G1, IsaacLab AMP formatember-lab-berkeley/AMASS_Retargeted_for_G1All of AMASS on G1, ready for IsaacLab's humanoid AMP taskUnitree G1Unitree G1 →cc-by-4.0Card license; built from fleaven/Retargeted_AMASS_for_robotics, so AMASS's non-commercial terms still apply upstream.IsaacLab AMP motion-loader format···3,024Hugging Face ↗
Retargeted AMASS for Fourier N1fleaven/Retargeted_AMASS_for_FourierN1AMASS motions retargeted to Fourier N1Fourier N1 (GRMini1T2 URDF)Fourier N1 →cc-by-4.0Card license; AMASS's original non-commercial terms apply to the source motion.Per-clip arrays [-1, 30]: root position, root quaternion, 23 joint positions···10,975Hugging Face ↗

Simulation and benchmarks · 2

Simulation and benchmarks datasets
DatasetEmbodimentLicenseFormatEpisodesFramesfpsHF dl 30dLinks
GR00T X-Embodiment Sim (GR1 humanoid subsets)nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-SimSimulated humanoid tabletop pick-and-place used to post-train GR00T N1Fourier GR1, simulated (gr1_arms_waist, gr1_unified, gr1_full_upper_body, gr1_arms_only)Fourier GR-1 →cc-by-4.0Per the dataset card···1,293,320Hugging Face ↗repo ↗
HumanoidBench27 tasks: 15 whole-body manipulation, 12 locomotionUnitree H1 with hands, simulated in MuJoCoUnitree H1 →MITThe repo's LICENSE file opens with MIT (UC Berkeley RLL) and also carries Apache-2.0 text; GitHub's detector reports NOASSERTION.Benchmark environments (code), not recorded data····humanoid-bench.github.io ↗repo ↗

How to read this

  • License is the dataset's own. CC BY-NC-SA data cannot train a model you sell. Gated datasets need a click-through on Hugging Face before download.
  • Motion capture is human data. AMASS and LAFAN1 are non-commercial, and a retargeted copy does not shed the source's terms, whatever license its own card shows. Retargets are kinematic: a policy still has to learn to track them on the real robot's dynamics.
  • HF downloads are the Hub's rolling 30-day count, fetched once a day. They measure use, not quality.
  • Missing a dataset? It needs a public host, a stated license and a humanoid you can build or buy. Then a policy trained on it can land in the policy zoo.