Steps of the work
Each step of a task is marked with its start, its end, who did it and whether it worked, with a reason when it failed.
Updates from skilled work & robotics.
Sign upCuratrix turns raw video of skilled work into a labelled, checked and privacy-safe dataset your team can train on.
Example of the clip editor: a class list, five objects on a frame from a Curatrix recording, the selected object's attributes and its track with keyframes on the timeline.
People mark what happens in each video, frame by frame, so your model learns from exact examples.
Shapes: boxes, rotated boxes, polygons, lines, keypoints with skeletons and brush masks. Tracks have keyframes and are interpolated in between; attributes can change per frame. Pre-labels come from detection and tracking models and stay marked as machine output until a person edits them.
What you see: a clip being labelled, with the list of things to mark, the video frame with its boxes and the timeline of the selected object.
Example interface, drawn from the Curatrix application. Clip names and content are illustrative.
Example of the label schema list with draft and published versions, and the editor of one class with its shape, parent, colour, hotkey and an attribute.
You decide what counts as a tool, a hand or a step once, and every labeller works from the same list.
A label schema defines classes, their shape type, colour, hotkey and attributes. Drafts can be saved freely; published versions are read-only, and each release pins the version it was labelled with. Fourteen sector presets ship as ontologies.
What you see: the list of label sets with their versions, and one set being edited, here the tools used on a building site.
Example interface, drawn from the Curatrix application. Clip names and content are illustrative.
Example of the judge view: a frame with solid and dashed boxes from two passes, a list of three conflicts and the verdict form with a required rationale.
Every clip is checked, a share of clips is labelled again by someone else, and disagreements are settled by a third person who writes down why.
Review policy per schema: single pass, second pass and judge on every clip, or adaptive second pass, with second-pass rate, blind-review rate, an escalation threshold on Cohen's kappa and a box-overlap (IoU) match threshold. Annotators qualify per sector on ten gold frames; the first tasks after qualifying get a second review.
What you see: the referee's view. Two versions of the same frame are compared, the three disagreements are listed, and each needs a decision and a reason.
Example interface, drawn from the Curatrix application. Clip names and content are illustrative.
Example of the face review: a video frame, three face regions on a timeline, two detected and one drawn by hand, and the sign-off button.
Faces in the footage are found, checked by a person and pixelated, and no dataset leaves until that check is signed off.
Face detection marks face regions; they are corrected on a timeline, signed off by an organisation admin and rendered into a pixelated copy. A release fails until every video is signed off and its render is newer than its last edit. Number plates, badges, screens and voices are not redacted.
What you see: the face check for one clip. The bars show when each face is on screen, and the button at the bottom signs the clip off.
Example interface, drawn from the Curatrix application. Clip names and content are illustrative.
Example of the dataset quality report: the release line, the episode table per project and the agreement metrics, with the figures left out.
Each dataset comes with a report that shows how the data was collected and checked, frozen at the moment of delivery.
Quality report version 4, pinned at release: episode counts per project, agreed acceptance criteria (passed, failed, unmeasured), protocol coverage, second reviews and agreement between passes as Cohen's kappa for labels, IDF1 for track identity and Krippendorff's alpha for action boundaries, each with median and 10th percentile.
What you see: the quality report of one dataset. The values are left empty here because every report takes its figures from its own delivery.
Example interface, drawn from the Curatrix application. Figures are left out because every report takes them from its own release.
Example of a released dataset: its version, the release manifest with passed checks, the archive and LeRobot downloads with checksums, and the files in the archive.
Every delivery is a numbered version with a digital signature, so your team can prove that nothing was changed or lost on the way.
The manifest lists the SHA-256 checksum of every file and is signed with Ed25519. Every release ships the main archive and a LeRobot v3.0 export; label packs in COCO, YOLO, CVAT XML, KITTI, RLDS and Hugging Face format are built with the release and chosen at download.
What you see: a delivered dataset with its version, the checks it passed, the download buttons and the files inside the archive.
Example interface, drawn from the Curatrix application. Clip names and content are illustrative.
The parts that hold a project together, from the first recording to the delivered dataset.
Each step of a task is marked with its start, its end, who did it and whether it worked, with a reason when it failed.
Every recording is checked before it is used: hands visible, camera steady, enough light, all steps there and sound recorded.
Only labellers who passed the test for your industry get your work, and their results stay on record.
Role changes, privacy sign-offs, deliveries and downloads of original footage are logged and cannot be edited.
Each organisation sees only its own data. Administrators sign off and deliver. Your team can sign in with its company login.
Your engineers can fetch and confirm deliveries automatically through an interface and a Python library.
See how it is set up for each industry and how it connects to your tools.
We walk you through the platform on footage from your field.