Note

This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.

Hugging Face

LeRobot Community Dataset v3 — Cross-Embodiment Starter#

image/png

A curated cross-embodiment starter subset drawn from lerobot/community_dataset_v3: 497 episodes across 50 unique robot embodiments, with up to 10 episodes from each embodiment.

Usage#

Load in FiftyOne#

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub("Voxel51/community_v3_10per_embodiment")
print(dataset)

# Filter by embodiment
so100_view = dataset.match(fo.ViewField("robot_type") == "so100")

Browse by robot type#

session = fo.launch_app(dataset)

# In App: filter sidebar by robot_type field to explore embodiments

Overview#

Metric

Value

Total Episodes

497

Robot Embodiments

50

FPS values

10, 15, 20, 25, 30, 50

Duration range

0.02 – 148.6 seconds

Source dataset

lerobot/community_dataset_v3 (v3.0)

Format

FiftyOne dataset (multimodal, LeRobotEpisodeReference)

Robot Types Covered (50 embodiments)#

Robot Type

Episodes

Source Dataset

Unitree_G1_Dex3

10

kuehnrobin/g1_stack_cube_left

Unitree_G1_Gripper

10

unitreerobotics/G1_MountCameraRedGripper_Dataset

Unitree_G1_Inspire

10

eastflag/g1_can_twohand

Unitree_G1_inspire

7

ccccccww/test1 (only 7 available total)

Unitree_Z1_Dual

10

unitreerobotics/Z1_DualArmStackBox_Dataset

Unitree_Z1_Single

10

unitreerobotics/Z1_StackBox_Dataset

agilex

10

Pi-robot/place_white_bottle

aiworker

10

noisyduck/ffw_bg2_rev4_ai_worker_demo_250714_2

aloha

10

HuaihaiLyu/test_shorts

aloha_solo

10

AshtinDonuts/aloha_test6

arx5

10

villekuosmanen/pick_coffee_capsule_under_dome

arx5_bimanual

10

villekuosmanen/move_glass_cups

bi_so100_follower

10

LeRobot-worldwide-hackathon/46-3580-bi_so100_pour_water_real

bi_xarm6_follower

10

kumarhans/ethernetGrab8

ffw

10

lcwoo/ffw_test

g1

10

hainh22/pick_cube_train

koch

10

Dongkkka/blue_pen_pick_and_place

koch_follower

10

aidinism/arm-record-14

mcx

10

relaxedandcalm/ds_move_the_motor

moss

10

Beegbrain/moss_stack_cubes

multi_robot

10

arclabmit/lx7r_knockover_test_dataset

mycobot320

10

GreenIsGood/pick_scan_and_place_warehouse_boxes_05_06_01_smthng

panda_follower

10

konstantinZ/pickup-single-cotton-V2

piper

10

hangwu/piper_joint_ep_20250421_realsense

piper_follower

10

CnLori/so101_piper

rx200

10

underctrl/toy_pickup_and_place

sam_bimanual

10

girardijp/sam_fold_cloth_single

sam_two

10

1g0rrr/demo2_upart_peeloff

so100

10

Gano007/so100_lolo

so100-blue

10

Chojins/chess_game_001_blue_stereo_flip

so100-red

10

Chojins/chess_game_001_red_stereo

so100MovellaDot

10

sergiov2000/so100_movella_pick_ball_C3 (5) + D3 (5)

so100_bimanual

10

Claessens14/playing_with_fold_april21_ego_single_fold_v1

so100_follower

10

triton7777/cube_to_bowl_so100

so100_follower_bimanual

10

LeRobot-worldwide-hackathon/160-LegoBot-lego_pixelart_stage_2_v2

so100_with_koch

10

rgarreta/so100_dataset1

so101

10

hawnsoung/so101_test_coff_5

so101_follower

10

MostafaOthman/pickplace_sweetsjp_smallcandy

stretch

10

Suzumiya894/stretch_test

stretch3

10

Suzumiya894/stretch_smolvla_dataset

trossen_ai_mobile

10

mrrl-emcnei/trossen_dataset0

trossen_ai_solo

10

minjunkevink/trossen_objects_pick_place

trossen_ai_stationary

10

ANRedlich/trossen_ai_stationary_test_08

unknown

10

easonjcc/rm_test_02

ur5e_gello

10

Arururu12/UR5e_Gello_Cube_bin_v2

widowx

10

jesbu1/usc_widowx_combined_play_data

xarm

10

Anas0711/orange_bobo_pose_action_state_june3_1130am

xarm6

10

kumarhans/glue_stick

xarm_end_effector

10

lukicdarkoo/pick_plazma

xarm_lite6

10

ReubenLim/xarm_lite6_gripper_cube_in_box

Dataset Structure#

Each sample in this FiftyOne dataset is one episode, with:

Field

Type

Description

robot_type

string

Embodiment identifier (50 unique values)

task

string

Task description

tasks

list[string]

All task descriptions for episode

episode_index

int

Episode index within the source dataset

length

int

Number of frames

duration

float

Duration in seconds

fps

float

Recording frame rate

media_reference

LeRobotEpisodeReference

Reference to video clips

dataset_name

string

Source contributor dataset (<contributor>/<dataset>)

qwen3vl_embedding

vector (2048)

Qwen3-VL-Embedding-2B video embedding (via qwen3vl_embeddings): up to 32 evenly spaced frames from the episode’s time window, last-token pooled and L2-normalised, in the same space as text embeddings from the same model

qwen3vl_embedding_camera

string

Camera stream used for qwen3vl_embedding

uniqueness

float

FiftyOne Brain uniqueness, computed from qwen3vl_embedding

representativeness

float

FiftyOne Brain representativeness, computed from qwen3vl_embedding

qwen3vl_embedding is a video embedding: each vector is computed from a clip spanning the episode, never from a single frame. The brain runs qwen3vl_embedding_sim (similarity index) and qwen3vl_embedding_umap (2D UMAP) are both built on it, and the similarity index accepts text queries as well as episodes. For text queries, register the model source first:

import fiftyone.zoo as foz

foz.register_zoo_model_source(
    "https://github.com/harpreetsahota204/qwen3vl_embeddings",
    overwrite=True,
)

view = dataset.sort_by_similarity(
    "folding a cloth", k=10, brain_key="qwen3vl_embedding_sim"
)

Parsing Decisions#

  • Episode selection: shard-0 episodes (first N from data/chunk-000/file-000.parquet and corresponding video chunk-000/file-000.mp4) for each source dataset. This gives the cheapest complete subset without splitting episodes across shards.

  • Cap: 10 episodes per embodiment, or all available if fewer than 10 exist.

  • Unitree_G1_inspire: Kept separate from Unitree_G1_Inspire; only 7 episodes existed in total across all contributing datasets — all 7 are included.

  • so100MovellaDot: No single source dataset had ≥10 episodes (max 5 per dataset). Combined 5 episodes from sergiov2000/so100_movella_pick_ball_C3 and 5 from so100_movella_pick_ball_D3 to reach 10.

  • Schema heterogeneity: The 50 embodiments use different camera setups, action dimensions, and FPS values. This FiftyOne dataset stores each sample document-style with only common fields (robot_type, task, duration, fps, length, episode_index). Camera-specific features are accessible via media_reference pointing to the source video clips.

  • No v2.x conversion: All 51 source sub-datasets were already at codebase_version v3.0 — no conversion needed.

  • No media re-encoding: Source video files are referenced as-is. Codecs may vary by embodiment (H.264, H.265, etc.).

  • episode_index values: Preserved from source datasets. Two so100MovellaDot sources both have indices 0–4; this is intentional — the media_reference key distinguishes them.

Source#

All data originates from lerobot/community_dataset_v3, a crowdsourced collection from 235 community contributors worldwide. Please acknowledge the original contributors when using this dataset.

License#

Apache 2.0 (same as source dataset). Individual datasets may have additional attribution requirements — see lerobot/community_dataset_v3 for contributor details.