annotation studio

Multimodal annotation, with an agent watching every frame.

One studio for 2D, 3D, LiDAR, and video. Annotators get suggested labels and next-step guidance as they work — each with a confidence score, a reason, and a required human sign-off.

The Annotation Studio is where the agentic feedback loop is most visible. As annotators draw boxes, cuboids, segments, keypoints, or event boundaries, agents compare the work against the guideline, the surrounding frames, and the rest of the dataset. When something looks off — a track that should have ended, an episode without an outcome, two events merged into one — the studio surfaces a specific suggestion rather than a generic warning. Every suggestion is applied only when a person accepts it.

2D and 3D labeling

Bounding boxes, polygons, semantic and instance segmentation, 3D cuboids, and keypoints with cross-view projection.

LiDAR point clouds

Point-level and cuboid labeling across sweeps, with ground removal, intensity coloring, and camera-fused views.

Video and event timelines

Frame-accurate tracks, sub-task segmentation, and outcome marking for long-horizon manipulation episodes.

Next-step recommendations

Split events, add boundaries, mark outcomes, or fix a track — suggested with confidence and one-click acceptance.

Provenance on every label

See what evidence produced a suggestion, who accepted or changed it, and which guideline version applied.

Guidelines that learn

Reviewer corrections are clustered into guideline clarifications, so the same ambiguity is resolved once.

how it works

Step by step

1

Open a task

The studio loads synchronized sensors for the episode with the relevant guideline pinned.

2

Annotate with assistance

Pre-labels and live suggestions appear inline; the annotator accepts, edits, or dismisses each.

3

Resolve flagged items

Ambiguous frames are routed to expert review with the full context attached.

4

Ship the batch

Accepted labels export with provenance, confidence, and guideline version.

Common questions

Do suggestions ever apply automatically?

No. Pre-labels can be shown, but nothing enters the dataset without a person confirming it.

Can we use our own labeling ontology?

Yes. Ontologies are versioned and shared across collection, annotation, review, and evaluation.

How are annotators onboarded?

With guided training tasks, gold-standard checks, and agent feedback calibrated to your guideline.

See it running on your data.

Start with a walkthrough on your own data.