Annotated Data and Delivery
In NexCore, annotated data is data that has been curated and processed by pipelines and carries structured annotations. It can be used as training input or downloaded as a standalone deliverable.
What Annotated Data Contains
| Annotation type | Source | Example |
|---|---|---|
| Quality labels | Manual labeling in the data workbench, or suggestions from batch-processing scripts | Anomaly_black frames, Anomaly_trajectory drift |
| Time ranges | Defined in the data workbench | Keep ranges, excluded ranges |
| Review status | Human review in the data workbench | Reviewed / Pending Review |
| Trajectories and poses | SLAM operators | 6-DoF camera trajectory |
| Depth | Depth operators | Per-frame depth maps |
| Structured visual results | Vision operators | Hand keypoints, detection boxes, segmentation masks |
| Robot state and actions | Included with teleoperation data | Joint positions, end-effector poses, gripper opening |
Generating Annotated Data
Ingest data → Create curation project → (RAW) Submit processing job → Review outputs → Label and review quality → Export dataset version| Step | Reference |
|---|---|
| Create a curation project | Curation Projects |
| Submit a processing job | Running Processing Jobs |
| Label and review quality | Data Workbench |
| Export | Datasets and Versions |
Reviewing Annotation Quality
Review each item in the data workbench, focusing on:
| Check | Looks good when | Common issues |
|---|---|---|
| Trajectory | Consistent with the motion in the video, with sensible start and end alignment | Drift caused by fast motion or low-texture scenes |
| Depth | Sharp object edges and no large holes | Gaps on reflective or transparent objects |
| Keypoints | Track the hands steadily | Lost when hands are occluded or leave the frame |
| Time alignment | All streams show the same event at the same moment | One channel lags behind |
| Ranges | Excluded ranges cover every failure segment | Range boundaries too wide or too narrow |
Label problem segments as anomalies and mark them as excluded ranges. If overall quality falls short, adjust the operator parameters and resubmit the processing job.
Export Formats
| Format | Use with |
|---|---|
| HDF5 | One file per sample; suited to in-house tools and training on the platform |
| LeRobot | Parquet + MP4; suited to popular open-source training frameworks |
Every dataset version comes with an export manifest that records each sample's data ID, source, review status, labels, and ranges. See Data Formats for the layout.
Downloading Annotated Data
Where: Data Curation → Dataset Management.
- Find a dataset version with status Exported.
- Click Download. The platform provides the data files and export manifest as a multi-part archive.
- Extract the archive and read the files as described in Data Formats.
For large datasets, the parts can be downloaded in parallel, and if one part fails, you only need to download that part again.
Delivering Through Delivery Orders
To formally deliver annotated data to a customer or partner, use a delivery order:
- Go to Data Delivery, click Create Delivery Order, select the dataset versions to deliver, and enter the recipient and a description.
- Click Check for Duplicates. The platform compares this delivery against previous deliveries to the same recipient and lists any duplicate samples.
- After submission, the order enters Pending Review and is reviewed by a member with approval permission.
- Once approved, the platform packages the data into a multi-part archive and publishes it to object storage. The order status changes to Published.
- The recipient downloads the data from the delivery link. You can bind the order to the recipient's email address.
| Order status | Meaning |
|---|---|
| Pending Review | Waiting for review |
| Queued / Processing | Packaging and publishing in progress |
| Published | Ready to download |
| Partially Succeeded / Failed | Some or all items failed to package or publish. Click Retry to retry only the failed items. |
| Rejected / Cancelled | Rejected at review, or cancelled while pending review, queued, or processing |
Why duplicate checks matter
The same samples are never delivered twice to the same recipient, which keeps delivery volumes easy to reconcile.
Common Issues
| Issue | Resolution |
|---|---|
| No Download button on a dataset | The dataset must be Exported, and your role must have export permission |
| A delivery order partially succeeded | Open the order details to see the failed items, then click Retry |
| Labels in the export manifest don't match expectations | The export reflects the project as it was at export time; after making changes, export a new version |

