01
Sources
Recursively discover and validate LeRobot datasets — Parquet, metadata, cameras, episode duration.
wrist
front
TWO-CAMERA CONTRACT
One canonical output, many sources
Every exported dataset carries a fixed wrist view plus a selectable second camera — rebuilt to a single consistent encoding, no matter which source it came from.
VALIDATION-FIRST
Nothing enters silently
Datasets are validated before use; preflight blocks export on any incompatibility.
NON-DESTRUCTIVE
Sources stay read-only
Curation, editing, and export never modify an original source dataset.
DETERMINISTIC
Balanced, reproducible
Per-prompt and per-group caps spread approved checkpoints evenly across tasks.
VERIFIED HANDOFF
v2.1 output
Rebuilt indices, edited prompts, and optional tarball packaging for handoff.
02
Output contract
Define the folder name, destination, camera pair, and optional per-task caps before curating.
03
Cameras & joints
Map source views to canonical cameras and auto-fill the six-joint SO-101 action/state order.
04
Episodes
Gallery with staging and all-camera preview, plus per-episode final prompts and flags.
05
Balance
Deterministic per-prompt and per-group caps with optional Groq-suggested group names.
06
Preflight & export
Blocking validation, then background jobs write normalized v2.1 data with rebuilt indices.
STATIC SNAPSHOT
The assembly workflow, explained
The demo is a faithful, static copy of the Dataset Assembly Studio interface. It walks all nine phases — Sources → Output → Cameras → Joints → Episodes → Tasks → Balance → Preflight → Export — describing what each step does.
SO-101 · FIXED WRIST
Curate
Review episodes, stage candidates, and set final prompts per episode or across the whole stage.
SO-101 · TWO CAMERAS
Balance
Keep the combined export evenly spread across prompts and task groups, with every checkpoint accounted for.
LEROBOT V2.1
Export
Background jobs write normalized data with rebuilt indices and edited prompts, optionally as a tarball.
STATIC DEMO
Explore the assembly workflow
No backend, no live data — just a walkthrough of how Dataset Assembly Studio takes raw sources to a verified LeRobot v2.1 export.
Launch the snapshot