Data Workbench
The data workbench is where you work on a curation project. It provides non-destructive curation and trimming: curation produces only metadata and export instructions, and source data stays read-only throughout.
Where: Data Curation → Data Workbench in the left navigation. Switch between curation projects at the top of the page.
Four-Step Workflow
Select data → Process (label / define ranges / review) → Autosave → ExportThe steps are fixed, so results don't depend on individual habits.
Page Layout
| Area | Contents |
|---|---|
| Header | Project switcher, new-data notice, Pipeline button, Export button, and a status bar while a pipeline is running |
| Left panel: data list | All items in the project, each showing data ID, duration, frame count, and label badges, with select-all and filters |
| Center: preview | Synchronized playback of all video streams, with the timeline and range tracks below |
| Bottom: charts | Joint angles, gripper state, and other robot-state curves, synced to the current frame; toggle individual joints |
| Right panel: annotations | Anomaly labels, ranges, review status, and actions for the current item |
Preview and Playback
- All camera streams play in sync by timestamp, at 0.5×, 1×, or 2× speed.
- Dragging the playhead moves both video and charts to the same frame.
- Keyboard shortcuts:
| Key | Action |
|---|---|
Space | Play / pause |
← / → | Back / forward one frame |
Shift + ← / Shift + → | Back / forward one second |
I / O | Set the current frame as the selection start / end |
E | Mark the current selection as an excluded range |
M | Add an anomaly label to the current selection |
R | Mark the current item as reviewed |
Quality Labels
All quality labels use the type Anomaly, with a note describing the specific issue. A label appears as Anomaly_<note>.
- Drag on the timeline to select a range, or press
I/Oto set the start and end frames. - Click Add Label and describe the issue in the note, for example "black frames", "dropped frames", or "missing joint data". A note is required.
- The label appears as
Anomaly_black frames, and a badge is shown in the data list.
- An item can have multiple labels, and their ranges can overlap.
- "Black frames", "dropped frames", and "missing joint data" are only common examples, not a fixed list. Describe what you actually observe.
Ranges: Keep and Exclude
You can define keep ranges and excluded ranges on the timeline:
| Action | Effect at export | Source file |
|---|---|---|
| Excluded range | That time span is trimmed out | Unchanged |
| Keep range | Only that time span is exported | Unchanged |
| Exclude episode | The entire item is left out of the export | Unchanged |
The system records only instructions. The export service generates the target data from those instructions, and the source file's hash never changes.
See it for yourself
After deleting a segment in the workbench, go back to Data Assets and open the source data. It is still there and plays back in full.
Human Review
Labeling and reviewing are different things. A label describes a quality issue; a review confirms that a person has inspected the item. Adding a label doesn't mark an item as reviewed.
- After inspecting an item, click Mark as Reviewed (shortcut
R). The reviewer and time are recorded. - Filter the data list by Pending Review or Reviewed to confirm nothing has been missed.
Batch Processing
Batch processing runs detection scripts across the whole project or a selection, producing candidate labels and processing suggestions:
- Click Batch Processing and choose a scope: all data, new data only, or a custom selection.
- Choose a mode: Incremental (only items not yet processed) or Full.
- Select detection scripts, such as black-frame detection, joint-data validation, or frame-rate anomaly detection. R&D engineers can upload custom scripts.
- Click Start Processing. The job runs asynchronously in the background.
| Result | Meaning |
|---|---|
| Completed | Every item was processed successfully |
| Completed with Errors | At least one item failed; the rest were processed normally. Check the logs to find the failed items, then retry or review them manually. |
| Failed | The job couldn't run. Check the logs and retry. |
Scripts never modify data
Scripts only produce candidate results, label suggestions, or processing steps. They can't silently delete or modify source data, and candidate labels take effect only after a person confirms them.
Reusable Processing Steps
Processing steps in the workbench (labels, ranges, reviews, and script runs) are saved automatically as step records that can be reused in other projects. This is the foundation for collaboration and iterative processing.
Export
When curation is complete, click Export in the header to create a dataset version. See Datasets and Versions. Projects can't be exported while a pipeline is running.

