A mobile robot tidying books and remotes into a basket on a coffee table
Home / Data annotation

Labels a policy can learn from.

Robotics labels live in 3D and in time. Trained annotators deliver segmentation, pose, grasp, action, and language labels through a two-pass QA loop — signal, not noise.

3D · POSE · GRASP
mug · 0.98 grasp_pt
A living-room scene with furniture labeled by colored bounding boxes, class names, and a point of interest marker
SPATIAL LABELS

Geometry, grounded.

2D/3D segmentation, 6-DoF pose, grasp points, and keypoints — labeled against calibrated multi-view footage and point clouds so they stay consistent across cameras and frames.

  • 2D / 3D segmentation
  • 6-DoF pose
  • grasp points
  • keypoints & skeletons
  • point-cloud labeling
TEMPORAL & ACTION LABELS

Structure in time.

Demonstrations segmented into skills and sub-tasks, with success/failure, recovery events, and contact phases marked frame-accurately — the structure skill-conditioned policies need.

  • skill segmentation
  • success / failure tags
  • contact-phase marks
  • recovery events
  • frame-accurate bounds
VISION · LANGUAGE · ACTION Two robot hands folding a white shirt on a table next to a laundry basket — a clear instruction-and-action pair
VLA PAIRS

Language that matches the motion.

We write and verify instructions, sub-goal captions, and paraphrase sets per episode — checked against the footage so language never drifts from what the robot did.

  • task instructions
  • sub-goal captions
  • paraphrase sets
  • hard negatives
  • multilingual on request
Quality loop

Every label passes through QA twice.

Annotation without measurement is guesswork. Each batch runs through a calibrated review loop before it ships.

01

Calibrate

Gold tasks and guideline tests qualify every annotator on your ontology first.

02

Annotate

Production labeling with embedded gold tasks tracking live accuracy per annotator.

03

Review

A second-pass reviewer audits geometry, timing, and language on every episode.

04

Score & ship

Batches ship with agreement stats, error taxonomies, and per-episode quality scores.

2×
Review passes per episode
98%+
Target label accuracy
100%
Episodes quality-scored
48h
Sample-batch turnaround
Try the suite

Send us a sample batch.

Share 20 episodes and your ontology. We'll return them fully labeled with QA metrics so you can judge the quality yourself.

Request sample labeling → or email contact@intellisant-ai.com