Skip to content

Curation Projects ​

A curation project is an editable workspace. It references read-only source data in the asset pool, records quality labels, ranges, and processing steps for that data, and is eventually exported as a dataset version.

Where: Data Curation → Curation Projects in the left navigation, or select data in Data Assets and click Add to Curation.

Curation Projects vs. Datasets ​

Curation projectDataset
NatureAn editable workspaceA fixed snapshot exported from a project
Editable?Yes. Add or remove data and change labels at any time.No. Frozen on export.
PurposeCuration and processingInput for training
Effect on source dataNone: references only, no copies or changesNone: exported files are stored separately

Creating a Curation Project ​

  1. Click New Curation Project to open the dialog.
  2. Enter a project name (required) and optional notes.
  3. Under Source Data, select one or more collection tasks as the candidate pool.
  4. Add filters as needed: device, format, scene, curation status, collection time, and so on.
  5. Select the items to include from the candidate list, one at a time or in bulk. The bottom of the dialog shows the matched count and the selected count in real time.
  6. Create becomes available once at least one item is selected. Click it, and the project opens in the data workbench.

Filtering is not the same as including

Filters only narrow down the candidates; items must be explicitly selected to be included. The platform deliberately has no option to add every item from a source in one click, to prevent accidental inclusion.

Data Rules Within a Project ​

  • Items are deduplicated by data ID within a project; adding the same item twice doesn't create a second row.
  • The same item can appear in multiple projects, and labels in one project don't affect the others.
  • A project should contain a single data type. HDF5 (teleoperation) data and RAW pipeline outputs go through different export pipelines, so don't mix them in one project.
  • Removing an item from a project deletes only its reference and labels within that project; the source data is unaffected.

Autosave ​

Every change in the data workbench (labels, ranges, reviews, exclusions) saves automatically, so there is no Save button. The platform keeps the complete curation history and audit trail, including who made each change and when.

Project Versions ​

Each time a project is exported as a dataset, the project version increases by one. The dataset records the project name and version at export time, so you can always tell which state of the project a dataset came from.

Collaboration ​

When several people edit the same project, the platform never silently overwrites anyone's changes, and every label records who made it. We recommend splitting the work by data scope, for example one person per collection task.

Next Steps ​

NexCore is evolving rapidly; the latest version prevails.NexCore Docs · Embodied AI Skill Learning Platform by Lumos Robotics