Data formats

Language / 语言： English | 简体中文

v2.1 and v3.0 are fully supported. Newer minor versions in the v2 and v3 families may open in read-only mode. Other versions, including v2.0, missing versions, and future major versions, are rejected.

---

Versions

codebase_version	Features	Export
v2.1	View and edit	Verified with official LeRobot v0.6.1
v3.0	View and edit	Same
Newer v2 or v3 minor version	View only	Not available
v2.0, other major versions, or missing version	Not opened	Not available

More runtime notes: Compatibility.

---

Inputs

Input	Notes
Local archive	.zip, .tar, .tar.gz, .tgz
Local folder	Drag or select a folder in supported browsers. File System Access (restorable handles) needs HTTPS or localhost; on HTTP LAN origins Studio uses the browser file picker instead.
Remote archive	CORS GET; large files also need Range — CORS
Custom	Implement DataSource

Archives are treated as untrusted input. A failed load does not change the source files.

---

Export (standalone app only)

Export is available only in the hosted or self-hosted standalone application, not in the npm viewer API.

For v3, Parquet files roll by size and chunks_size caps files per chunk. Videos use one MP4 per episode. The exported layout is not byte-identical to the official sharding scheme, but it has been verified with the pinned LeRobot v0.6.1 reader.

Subtasks

Subtasks are optional. When present they follow the official LeRobot v3.0 mapping:

• meta/subtasks.parquet: pandas string index subtask plus subtask_index (int64)
• info.features.subtask_index: { "dtype": "int64", "shape": [1], "names": null }
• frame parquet column subtask_index

A dataset without subtasks is exported without these files or the feature. Checking Include subtasks on v3.0 export writes these files and requires complete frame coverage. Unchecked exports omit subtask files and skip that check. Export to v2.1 strips subtask columns and meta/subtasks.parquet. The original task / task_index fields are never overwritten.

Studio reads official subtask_index only. Community snapshots that store subtask_id instead are opened as ordinary v3.0 data, but subtask segments are not shown. Missing meta/subtasks.parquet is allowed: labels fall back to Subtask N. Official unlabeled frames may use subtask_index = -1; Studio treats those as unlabeled and never writes -1 on export. Extra official columns such as task_index_high_level are preserved. A v3.0 export with Include subtasks checked always writes meta/subtasks.parquet.

Official Hub examples

Download with the Hugging Face CLI (hf), then open the local folder in Studio:

hf download lerobot/libero_10_subtask --type=dataset --local-dir /data/lerobot/libero_10_subtask
hf download lerobot/pusht-subtask --type=dataset --local-dir /data/lerobot/pusht-subtask
hf download k1000dai/libero-subtaskid-segments --type=dataset --local-dir /data/lerobot/libero-subtaskid-segments
hf download lerobot/outdoor-depth --type=dataset --local-dir /data/lerobot/outdoor-depth

Hub repo	What it contains	Studio
lerobot/libero_10_subtask	Official v3.0 with frame subtask_index. Labels parquet may be absent.	View indices as Subtask N; annotate and export v3.0.
lerobot/pusht-subtask	Official v3.0 with meta/subtasks.parquet (phase 1/2/3) and unlabeled -1 frames.	View labels; unlabeled frames stay empty; including subtasks on export requires full coverage.
k1000dai/libero-subtaskid-segments	v3.0 snapshot that uses subtask_id instead of official subtask_index.	Opens for viewing; subtask segments are not mapped.
lerobot/outdoor-depth	Official v3.0 depth-only dataset. Frames are 16-bit grayscale TIFF in parquet (is_depth_map: true, depth_unit: mm). No subtasks.	Decodes TIFF depth maps and plays a false-color preview.