datasets.humanoidrobots.training · checked 2026-09-25
Open robot datasets
Data to train on, picked for robots you can actually build or buy. 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; blank means the source does not state it.
- Datasets
- 24
- LeRobot format
- 16
- Episodes (stated)
- 100,479
- Non-commercial
- 2
Low-cost arms · 5
Record your own with LeRobot
| Dataset | Embodiment | License | Format | Episodes | Frames | fps | HF dl 30d | Links |
|---|---|---|---|---|---|---|---|---|
| SmolVLA SO-101 pick-and-placelerobot/svla_so101_pickplacePick a cube and place it in a box | SO-101 follower arm (robot_type so100_follower)SO-101 (SO-ARM101) → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 50 | 11,939 | 30 | 7,931 | Hugging Face ↗ |
| SmolVLA SO-100 pick-and-placelerobot/svla_so100_pickplacePick a cube and place it in a box | SO-100 armSO-100 (SO-ARM100) → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 50 | 19,631 | 30 | 12,727 | Hugging Face ↗ |
| Koch pick-and-place, 5 Lego brickslerobot/koch_pick_place_5_legoPick five Lego bricks and place them | Koch v1.1 armKoch v1.1 Low-Cost Robot Arm → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 50 | 37,972 | 30 | 472 | Hugging Face ↗ |
| Koch pick-and-place, 1 Lego bricklerobot/koch_pick_place_1_legoPick one Lego brick and place it | Koch v1.1 armKoch v1.1 Low-Cost Robot Arm → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 102 | 39,693 | 30 | 364 | Hugging Face ↗ |
| ALOHA static: coffeelerobot/aloha_static_coffeeBimanual coffee making | ALOHA bimanual (ViperX arms)ALOHA → | mit | LeRobot v3.0 · parquet + mp4 | 50 | 55,000 | 50 | 3,056 | Hugging Face ↗ |
Humanoids · 10
| Dataset | Embodiment | License | Format | Episodes | Frames | fps | HF dl 30d | Links |
|---|---|---|---|---|---|---|---|---|
| Unitree G1 + Dex3: toasted breadunitreerobotics/G1_Dex3_ToastedBread_DatasetBimanual toast handling | Unitree G1 with Dex3 hands (robot_type Unitree_G1)Unitree G1 → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 418 | 352,022 | 30 | 6,840 | Hugging Face ↗ |
| Unitree G1 + Dex1: stack blocksunitreerobotics/G1_Dex1_Stack_BlockStack blocks | Unitree G1 with Dex1 grippersUnitree G1 → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 690 | 284,191 | 30 | 4,246 | Hugging Face ↗ |
| Unitree G1 + Dex1: fold towelunitreerobotics/G1_Dex1_Fold_TowelFold a towel | Unitree G1 with Dex1 grippers (robot_type G1_ALL)Unitree G1 → | apache-2.0 | LeRobot v2.1 · parquet + mp4 | 200 | 310,671 | 30 | 2,521 | Hugging Face ↗ |
| Unitree G1 + BrainCo hands: pick appleunitreerobotics/G1_Brainco_PickApple_DatasetPick an apple | Unitree G1 with BrainCo dexterous handsUnitree G1 → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 200 | 153,668 | 30 | 695 | Hugging Face ↗ |
| Unitree H1: fold clotheslerobot/unitreeh1_fold_clothesFold clothes | Unitree H1Unitree H1 → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 38 | 19,000 | 50 | 425 | Hugging Face ↗ |
| Unitree H1: warehouselerobot/unitreeh1_warehouseWarehouse pick and place | Unitree H1Unitree H1 → | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 24 | 11,275 | 50 | 472 | Hugging Face ↗ |
| Humanoid EverydayUSC-PSI-Lab/humanoid-everydayEveryday manipulation and loco-manipulation | Humanoid robots, per the card (arXiv 2510.08807) | apache-2.0 | Video + robot data, per the dataset card | · | · | · | 27,929 | Hugging Face ↗ |
| AgiBot World Alphaagibot-world/AgiBotWorld-AlphaReal-world dual-arm manipulation across many scenes | AgiBot dual-arm robots | CC BY-NC-SA 4.0gatedStated in the gated card's agreement and the repo README; the card has no license field. Non-commercial. | WebDataset | · | · | · | 30,629 | Hugging Face ↗agibot-world.com ↗repo ↗ |
| AgiBot World Betaagibot-world/AgiBotWorld-BetaThe larger follow-up release of AgiBot World | AgiBot dual-arm robots | CC BY-NC-SA 4.0gatedStated in the repo README for all data and code; the HF card has no license field. Non-commercial. | WebDataset | · | · | · | 85,355 | Hugging Face ↗agibot-world.com ↗repo ↗ |
| RoboMINDx-humanoid-robomind/RoboMINDMulti-embodiment manipulation | Multiple embodiments including humanoids, per the card | apache-2.0gated | Per the dataset card | · | · | · | 35,924 | Hugging Face ↗ |
Cross-embodiment, large · 5
| Dataset | Embodiment | License | Format | Episodes | Frames | fps | HF dl 30d | Links |
|---|---|---|---|---|---|---|---|---|
| DROID 1.0.1lerobot/droid_1.0.1In-the-wild manipulation, language-annotated | Franka Panda (robot_type Franka), many labs | apache-2.0 | LeRobot v3.0 · parquet + mp4 | 95,658 | 27,630,375 | 15 | 22,096 | Hugging Face ↗droid-dataset.github.io ↗ |
| Open X-Embodimentjxu124/OpenX-EmbodimentPooled manipulation data behind RT-X | 22 robot embodiments, per the project | cc-by-4.0License field of the Hugging Face mirror (a community mirror, not the official release). Component datasets carry their own terms; the official repo is Apache-2.0 code. | RLDS | · | · | · | 23,533 | Hugging Face ↗robotics-transformer-x.github.io ↗repo ↗ |
| Berkeley Autolab UR5 (Open X subset)lerobot/berkeley_autolab_ur5Tabletop manipulation | UR5 arm | cc-by-4.0 | LeRobot v3.0 · parquet + mp4 | 1,000 | 97,939 | 5 | 1,402 | Hugging Face ↗ |
| PhysicalAI GR00T X-Embodiment Simnvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-SimSimulated manipulation for post-training GR00T | Simulated embodiments used for GR00T N1, per the card | cc-by-4.0 | Per the dataset card | · | · | · | 1,070,809 | Hugging Face ↗ |
| BridgeData V2Tabletop manipulation across many environments | WidowX 250 arm | see project siteThe training code repo is MIT; data terms are on the project site. No Hugging Face copy verified. | Per the project site | · | · | · | · | rail-berkeley.github.io ↗repo ↗ |
Simulation and benchmarks · 4
| Dataset | Embodiment | License | Format | Episodes | Frames | fps | HF dl 30d | Links |
|---|---|---|---|---|---|---|---|---|
| ALOHA sim insertion (human demos)lerobot/aloha_sim_insertion_humanPeg insertion, teleoperated in sim | ALOHA bimanual, simulatedALOHA → | mit | LeRobot v3.0 · parquet + mp4 | 50 | 25,000 | 50 | 46,028 | Hugging Face ↗ |
| LIBEROlerobot/libero40 language-conditioned tasks, the standard VLA sim benchmark | Franka Panda, simulated (robot_type panda) | apache-2.0Card license of the LeRobot conversion; the LIBERO code repo is MIT. | LeRobot v3.0 · parquet + mp4 | 1,693 | 273,465 | 10 | 17,592 | Hugging Face ↗libero-project.github.io ↗repo ↗ |
| Push-Tlerobot/pushtPush a T block into a target pose (Diffusion Policy benchmark) | 2D point pusher, simulated | mit | LeRobot v3.0 · parquet + mp4 | 206 | 25,650 | 10 | 24,426 | Hugging Face ↗ |
| HumanoidBench27 whole-body locomotion and manipulation tasks | Unitree H1 with hands, simulated in MuJoCoUnitree H1 → | not statedGitHub reports no recognised license for the repo (NOASSERTION); check the repo before reuse. | 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.
- 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 robot you can build or buy. Then a policy trained on it can land in the policy zoo.