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.

This is a FiftyOne dataset with 369 samples.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

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

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/interndata-a1-close-the-laptop-369ep")

# Launch the App
session = fo.launch_app(dataset)

Dataset Card for InternData-A1 close-the-laptop#

Dataset Details#

Dataset Description#

This is a 369-episode subset of the franka-1/articulation_tasks/close_the_laptop task directory from griffinlabs/InternData-A1-LeRobot-v3.0-by-embodiment, itself a re-packaging of InternRobotics/InternData-A1 with tarballs extracted and the directory structure reorganized so the top-level directories are per-embodiment.

The source repo is not a single LeRobot dataset: it contains 1,488 independent LeRobot v3.0 datasets, one per <embodiment>/<task_category>/<task>[/<object_variant>] directory, each with its own meta/info.json, data/, and videos/. This subset covers one of those 1,488 task directories: franka-1 (a Franka arm using the franka-1 feature/gripper convention, one of two Franka variants in the source repo), articulation_tasks/close_the_laptop. Every episode in the “shard-0” set (episodes whose data shard and every video-stream shard are the first file, file_index == 0, of chunk-000) was imported — 369 of the task’s 578 total episodes, the largest complete subset obtainable without downloading every shard.

Curated by: Griffin Labs (source re-packaging); this subset selected and exported by a FiftyOne user via the fiftyone-lerobot-subset-to-hub skill
Language(s): English (task language instructions)
License: cc-by-nc-sa-4.0 (inherited from the source repo)

Dataset Sources#

Repository: https://huggingface.co/datasets/griffinlabs/InternData-A1-LeRobot-v3.0-by-embodiment
Original dataset: https://huggingface.co/datasets/InternRobotics/InternData-A1

Uses#

Direct Use#

Exploring, filtering, and visualizing Franka “close the laptop” articulation episodes in the FiftyOne App; prototyping multimodal (video + tabular state/action) data loaders; small-scale policy debugging or overfitting tests on a single task.

Out-of-Scope Use#

Not suitable for training a policy expecting the task’s full episode coverage — this is 369 of 578 total episodes available in the source repo for this task. Not suitable for commercial use without checking the CC BY-NC-SA 4.0 terms. Not representative of any embodiment or task other than a single Franka arm performing this one articulation task.

Dataset Structure#

Media type: multimodal. One sample = one episode. Each sample’s media_reference points to the episode’s rows in this self-contained, re-exported LeRobot v3.0 Parquet + MP4 export (re-indexed to episode_index 0..368).

Fields#

Field	Type	Meaning
`id`	ObjectIdField	FiftyOne sample id
`media_reference`	MediaReferenceField	Pointer to the episode’s data/video rows in the LeRobot export
`tags`	ListField(StringField)	FiftyOne tags (empty by default)
`metadata`	EmbeddedDocumentField	FiftyOne media metadata (unset)
`created_at` / `last_modified_at`	DateTimeField	FiftyOne bookkeeping
`episode_index`	IntField	Episode index within this exported dataset
`task`	StringField	Task instruction for the episode (“Close the laptop” for all samples)
`tasks`	ListField(StringField)	All task instruction strings associated with the episode
`length`	IntField	Number of frames in the episode
`duration`	FloatField	Episode duration in seconds (`length / fps`)
`robot_type`	StringField	`Franka` for all samples
`fps`	FloatField	Recording frame rate (30.0)

task is a StringField, not a Classification, because in LeRobot v3 the task instruction is a free-text string looked up per-frame via task_index, not a fixed label set; all 369 episodes in this subset share the single task instruction “Close the laptop”.

Per-frame streams (not sample-level fields)#

Per-frame video and tabular data live in the referenced Parquet/MP4 shards and appear in the FiftyOne App’s Streams / State & Action tabs, not as top-level sample fields:

Feature	dtype	shape
`images.rgb.head`	video	`[360, 640, 3]`
`images.rgb.hand`	video	`[480, 640, 3]`
`head_camera_intrinsics`	float32	`[4]`
`hand_camera_intrinsics`	float32	`[4]`
`head_camera_to_robot_extrinsics`	float32	`[7]`
`hand_camera_to_robot_extrinsics`	float32	`[7]`
`states.joint.position`, `actions.joint.position`, `master_actions.joint.position`	float32	`[7]`
`states.gripper.position`, `actions.gripper.position`	float32	`[1]`
`actions.gripper.openness`, `master_actions.gripper.position`, `master_actions.gripper.openness`	float32	`[1]`
`states.gripper.pose`, `actions.gripper.pose`, `master_actions.gripper.pose`	float32	`[6]`
`states.ee_to_armbase_pose`, `states.ee_to_robot_pose`, `states.tcp_to_armbase_pose`, `states.tcp_to_robot_pose`, `states.robot_to_env_pose` (and matching `actions.*`)	float32	`[7]`
`observation.state`	float32	`[49]`
`action`	float32	`[43]`

All declared features are video or standard scalar (float32/int64) dtypes, so nothing was excluded by the FiftyOne importer.

observation.state and action are derived columns, added after import by concatenating the states.* and actions.* components above (in the fixed order listed in the table) into single vectors, with a names array labeling each dimension (e.g. states.joint.position[0], …, states.gripper.position, …). FiftyOne’s App only renders its State & Action tile for features literally named observation.state / action; InternData-A1’s native per-component naming (states.joint.position, actions.gripper.pose, …) isn’t pattern-matched by that tile. The original per-component columns are left untouched alongside the two new vectors, so no information from the source is lost or duplicated-away. master_actions.* (the teleop leader-arm commands, distinct from the executed actions.*) was intentionally excluded from the action vector and is only available via its original per-component columns.

`dataset.info` contents#

ds.info["lerobot"]["skipped_episodes"] is [], confirming all 369 requested episodes imported successfully.

Parsing decisions#

Why this task: this repo was originally scoped as a 10-task, 5-embodiment “representative mix” spanning Franka, Genie-1, ARX Lift-2, and AgileX Split Aloha. Merging those into one Hub push turned out not to be possible: FiftyOne’s LeRobotDatasetExporter requires all exported episodes to share one source_id (raises MediaReferenceError: LeRobotDataset export cannot mix episodes from different sources), and even independent of that check, the LeRobot v3 format requires one fixed feature schema per dataset — the single-arm Franka schema (images.rgb.hand, states.joint.position, …) and the dual-arm Genie-1/Lift-2/ Split-Aloha schema (images.rgb.hand_left/hand_right, states.left_joint.position/right_joint.position, …) are incompatible within one info.json. This repo covers just one of the original 10 candidate tasks, franka-1/articulation_tasks/close_the_laptop, as a valid, self-contained example.
Shard-0, not “all episodes”: 369 of 578 total episodes for this task were imported — every episode whose data/file_index and every video stream’s file_index are 0 within chunk-000, the largest complete subset obtainable while downloading only the first Parquet/MP4 shard per stream.
Metadata repair applied: the source meta/tasks.parquet stored the task text as the pandas index (columns ['task_index', '__index_level_0__']) rather than a literal task column, which FiftyOne’s exporter requires (_read_source_tasks raised MalformedMediaSourceError). Repaired locally by rewriting the file with explicit task_index/task columns (df.reset_index().rename(columns={"index": "task"})), then re-importing this FiftyOne dataset from the repaired source before exporting.
No modalities excluded: the task declares only video and standard scalar dtypes, so nothing was dropped by the importer.
Re-export re-indexing: episode_index and task_index were local to the source task directory (0..577) before this export; the push remaps the 369 selected episodes to a contiguous 0..368 range and recomputes aggregate stats over just this subset.
License: unchanged from the source (cc-by-nc-sa-4.0, non-commercial). Per the source README, franka-1 and franka-2 (not used here) differ in image shape and gripper value range; this subset uses only franka-1.
observation.state/action packing: added post-export (see Fields above) so the FiftyOne App’s State & Action tile — which only recognizes those two literal feature names — renders this dataset’s proprioception. Verified the packed vectors exactly reproduce the original per-component values before pushing.

Dataset Creation#

Curation Rationale#

To provide a small, self-contained, browsable FiftyOne/LeRobot v3 export of a single Franka articulation task from the very large (2.1 TB, 1,488-task) InternData-A1- LeRobot-v3.0-by-embodiment collection.

Source Data#

Data Collection and Processing#

Per the source repo’s card, InternData-A1 was reorganized from InternRobotics/InternData-A1 by extracting the original tarballs and “transposing” the directory structure so the top-level directories are the embodiments. This FiftyOne subset further selected the franka-1/articulation_tasks/close_the_laptop task directory, imported its shard-0 episode set after repairing meta/tasks.parquet, and re-exported as a self-contained LeRobot v3.0 repository with episodes re-indexed to 0..368.

Annotations#

Annotation process#

The task instruction (“Close the laptop”) is a free-text string provided in the source meta/tasks.parquet; no additional annotation was added for this subset.

Dataset Card Authors#

Harpreet Sahota

Dataset Card Contact#