Annotation

Annotation for robot data.

Our in-house vision-language models are trained for accurate, cost-efficient robot annotation. QC and human review produce training-ready data.

Build with RoboLens ↗

Video, robot state and tactile signals.

  1. In-house VLMs

    Generate annotations to match your label definitions and output schema.

  2. Quality control

    Check consistency, completeness and alignment with your labeling guidelines.

  3. Human review

    Reviewer network across India, China and Latin America.

    Correct flagged annotations and spot-check quality against source data.

Training-ready data

Reviewed annotations with source references in your training format.

Illustrative workflow. Label definitions and review criteria are set for your project.

Egocentric, teleoperation and UMI data.

From pretraining to deployment.

Egocentric recording RGB video

Hand keypoints & action annotations

Hand pose aligned to each frame. Task steps aligned to time.

00:00 / 00:36
Source recording Raw video
Hand keypoints Projected 3D hand pose

Action annotations

Select a task step to inspect its interval.

Unlabeled interval

Robot teleoperation RGB video

Action annotations

Timestamped task steps from a human-operated robot.

00:00 / 00:25
Source recording Raw video

Action annotations

Select a task step to inspect its interval.

0.0–2.5 s

Autonomous robot run RGB video

Task outcome annotations

The observed result, checked against the task definition.

Evidence clip · source 9.0–15.0 s

Placement unsuccessful

Task
Put the red marker in the purple bowl
Observed
The red marker is released into the blue bowl; the task requests the purple bowl.
Labeled interval
12.4–13.0 s in the source recording

Robot manipulation Tactile + gripper signals

Contact event annotations

Annotate contact, grasp and slip events using tactile and gripper signals.

Touch at each finger

2 sensor counts29 sensor counts10 sensor counts6 sensor counts0 sensor counts41 sensor counts48 sensor counts17 sensor counts-11 sensor counts7 sensor counts23 sensor counts14 sensor counts-9 sensor counts-6 sensor counts7 sensor counts7 sensor counts-8 sensor counts-1 sensor counts-1 sensor counts3 sensor counts-3 sensor counts-10 sensor counts-1 sensor counts-3 sensor counts-2 sensor counts3 sensor counts9 sensor counts-2 sensor counts
Finger 1
5 sensor counts12 sensor counts12 sensor counts9 sensor counts-10 sensor counts16 sensor counts29 sensor counts12 sensor counts23 sensor counts32 sensor counts59 sensor counts28 sensor counts0 sensor counts6 sensor counts13 sensor counts13 sensor counts-3 sensor counts-5 sensor counts8 sensor counts3 sensor counts-5 sensor counts2 sensor counts-2 sensor counts4 sensor counts-8 sensor counts-3 sensor counts-2 sensor counts-5 sensor counts
Finger 2
-2976

Sensor counts · one scale for both fingers

Tactile response

Sum of 28 taxels · counts
347-179
Finger 1 169Finger 2 243

Gripper position

Position counts
1270
5.8 s8.3 s10.8 s

Measured position 127

Event annotations

8.3 s
  1. 6.4–7.7 sGripper movementSensor proposal
  2. 7.7–9.2 sHold contactDraft annotation
  3. 8.8–10.3 sSlipReviewed source label

Recorded sensor measurements. Each event retains its source review status.

Example label scopes

Define the labels around the task.

Set the action boundaries, outcomes and review criteria around your robot recordings.

Task execution

Action boundaries, task steps and outcomes.

Retries & interventions

Repeated attempts, recoveries and operator assistance.

Contact-rich tasks

Contact, release and slip events with sensor evidence.