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.

Dataset Card for SharpaDex Assemble Gears (one season, 410 episodes)#

SharpaDex Assemble Gears preview

This is a FiftyOne dataset with 410 samples. Each sample is one teleoperated bimanual episode of the task “Gear Assembly on Base”.

Installation#

pip install -U fiftyone

Usage#

Load from the Hub#

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

dataset = load_from_hub("Voxel51/sharpa-assemble-gears-410ep")

session = fo.launch_app(dataset)

Download a snapshot and load it locally#

import fiftyone as fo
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Voxel51/sharpa-assemble-gears-410ep",
    repo_type="dataset",
    local_dir="sharpa-assemble-gears-410ep",
)

dataset = fo.Dataset.from_dir(
    dataset_dir="sharpa-assemble-gears-410ep",
    dataset_type=fo.types.LeRobotDataset,
    name="sharpa_assemble_gears",
)

session = fo.launch_app(dataset)

Dataset Details#

Dataset Description#

One collection season (season_POC22027_2026_04_30_15_04_50_train) of the assemble_gears_on_base task from SharpaDex v1.0, taken from its lerobot_v3.0 export: 410 episodes, 302,020 frames at 30 fps. The full SharpaDex v1.0 release has 28,993 episodes across 59 tasks and 495 seasons (about 300 hours); this FiftyOne dataset covers only this one season of one task.

Curated by: Sharpa
Shared by: Sharpa-Robotics on Hugging Face
Language(s): en
License: CC-BY-4.0

Dataset Sources#

Repository: https://huggingface.co/datasets/Sharpa-Robotics/SharpaDex-v1.0 (path: assemble_gears_on_base/season_POC22027_2026_04_30_15_04_50_train/lerobot_v3.0)
Demo: https://www.sharpa.com/

Uses#

Direct Use#

Per the source README: research on imitation learning, visuomotor control, vision-language-action models, visual-tactile learning, and hierarchical policy learning.

Dataset Structure#

Topology. media_type: multimodal, 410 samples, one per episode, 30 fps, 523-1,101 frames (17.4-36.7 s), 302,020 frames total. Samples hold a media_reference to the per-frame parquet data and 6 video streams.

Field	FiftyOne type	Description
`media_reference`	`LeRobotEpisodeReference`	Pointers into `data/` parquet and the 6 video streams
`episode_index`	`IntField`	Episode index (0-409)
`task`	`StringField`	Structured task text (`Task`, `Instruction`, `Scene`, `Success`)
`tasks`	`ListField(StringField)`	All task strings of the episode
`length`	`IntField`	Frames in the episode
`duration`	`FloatField`	Episode duration in seconds
`robot_type`	`StringField`	`robot` (the value declared in the source `info.json`)
`fps`	`FloatField`	30

Data behind media_reference (per-frame, from meta/info.json and the source README):

Feature	Shape	Description
`observation.state`	(65,)	Joint state: left arm 0-6, left hand 7-28, right arm 29-35, right hand 36-57, torso/motor joints 58-64
`action`	(65,)	Joint-space action target, same layout
`observation.state.joint_torque`	(65,)	Joint-torque signal, same layout
`observation.state.tcp`	(24,)	Left pose 0:6, left force/torque 6:12, right pose 12:18, right force/torque 18:24
`observation.state.tcp_pose`	(12,)	Observed TCP poses, left 0:6 and right 6:12
`action.tcp_pose`	(12,)	Commanded TCP poses, left 0:6 and right 6:12
`observation.tactile`	(60,)	10 fingertips x 6-axis force/torque signal
`subtask_index`	(1,) int64	Index into the source subtask table
`observation.images.head_left`, `head_right`, `wrist_left`, `wrist_right`	480x480x3	RGB cameras (H.264 yuv420p)
`observation.images.tactile_deform`	480x1200x3	Tactile deformation video (H.264 yuv420p)
`observation.images.tactile_raw`	480x1600x3	Raw tactile video (H.264 yuv420p)

Label types and why. The importer produces episode-level samples with a media_reference; numeric streams (state, action, torque, tactile force/torque, TCP) are read through the viewer and are not stored as FiftyOne label fields. The source’s temporally grounded language segments (start/end step, text, skill) are not imported as FiftyOne fields. No custom fields were added, so the FiftyOne dataset matches what the LeRobot export contains.

dataset.info. lerobot entry with format, format_major, episode_count, imported_episode_count, skipped_episodes (empty).

Parsing decisions.

One season was chosen as a bounded, self-contained slice (19.6 GB) of a 3.4 TB repo. It is a single contiguous season, not a random sample of the task or the release.
All 6 video streams were imported; no modality was excluded. No source file was modified.
The source README states the export does not fully declare TCP units, reference frames or rotation convention (rotation_type is null). Do not infer Euler angles or quaternions from the vector width alone.
robot_type is the literal string robot in the source info.json.
Sidecar metadata not carried by the export: the FiftyOne LeRobot exporter writes only info.json, stats.json, tasks.parquet, the episodes table, data/ and videos/. The source meta/subtasks.parquet, meta/annotations.jsonl and meta/modality.json are therefore not in the pushed repo. The subtask_index column is kept in the per-frame parquet, but the subtask text and skill labels it points to are available only in the source repo.

Dataset Creation#

Curation Rationale#

Per the source README: a real-world teleoperation dataset for bimanual dexterous manipulation, covering assembly, tool use, deformable-object manipulation, cleaning, material transfer, and long-horizon tasks.

Source Data#

Data Collection and Processing#

Per the source README: human teleoperation of a real bimanual robotic system with two 7-DoF arms, two 22-DoF dexterous hands, four RGB cameras, and fingertip tactile sensors. Each episode provides synchronized joint state, joint torque, tool-center-point state, observed and commanded TCP poses, action, numeric tactile measurements, RGB video, tactile video, and temporally aligned language annotations.

Who are the source data producers?#

Sharpa.

Annotations#

Annotation process#

Per the source README: every episode has a global structured task description (Task, Instruction, Scene, Success) and temporally grounded subtask segments with a skill label.

Citation#

BibTeX:

@misc{sharpadex_v1_2026,
  title        = {SharpaDex v1.0: Large-Scale Bimanual Dexterous Manipulation Demonstrations},
  author       = {{Sharpa}},
  howpublished = {\url{https://huggingface.co/datasets/Sharpa-Robotics/SharpaDex-v1.0}},
  year         = {2026}
}

APA:

Sharpa. (2026). SharpaDex v1.0: Large-Scale Bimanual Dexterous Manipulation Demonstrations. https://huggingface.co/datasets/Sharpa-Robotics/SharpaDex-v1.0