Every movement,
labeled.

Turn robot demonstrations into training data with step boundaries, instructions and human review.

Explore the live demo
326 real episodes processed in our demo batch•$0.02 measured AI drafting cost per episode•100% of episodes checked by a person•Audit trail on every label and every edit•LeRobot datasets supported today•
Episode review · 01 / 03

Was the towel placed?

Check where the towel lands after release. Record success, failure or an unclear outcome.

Outcome · reviewer records
successfailureunclear
Illustrated robot gripper holding a folded towel above a metal bowl
Outcome checkPlacement region
Illustrated review · Drag to compare

Why it's different

01 / 04

Annota fixes
step boundaries.

— Typical labeling todayAnnota
BoundariesScrubbing video by hand to find where each step starts.Cut from gripper and arm signals, to the frame.
EvidenceA timestamp in a spreadsheet, with no reason attached.Every boundary keeps the signal that produced it.
ReviewEvery episode relabeled from zero.A person fixes only what the signals got wrong.
See it on a real batch →

How it works

From raw episodes
to reviewed labels.

How it works

From raw episodes
to reviewed labels.

DROID / Episode 14AI draft
Wrist camera at 5.07 secondsWrist camera · 5.07 s

Reach → grasp

01 / 04

Find the grasp.

The arm reaches for the towel. At 5.07 seconds the gripper closes, ending the reach and starting the grasp.

02 / 04

Name the movement.

The vision model reads the next interval and drafts an instruction that describes what the robot does with the towel.

03 / 04

Check the release.

A reviewer can inspect the moment after release. Does the towel stay in the bowl as the gripper retreats? The answer helps determine the outcome.

04 / 04

Read the labeled episode.

Five intervals connect the footage to the instructions. Open the batch to inspect other episodes, their outcomes and quality flags.

01 / 04

Find the grasp.

The arm reaches for the towel. At 5.07 seconds the gripper closes, ending the reach and starting the grasp.

DROID / Episode 14AI draft
Wrist camera at 5.07 secondsWrist camera · 5.07 s

Reach → grasp

02 / 04

Name the movement.

The vision model reads the next interval and drafts an instruction that describes what the robot does with the towel.

DROID / Episode 14AI draft
Side camera at 8 secondsSide camera · 8.00 s

Carry the towel to the bowl.

03 / 04

Check the release.

A reviewer can inspect the moment after release. Does the towel stay in the bowl as the gripper retreats? The answer helps determine the outcome.

DROID / Episode 14AI draft
Front camera at 13.5 secondsFront camera · 13.50 s

Review question · Is the towel in the bowl?

04 / 04

Read the labeled episode.

Five intervals connect the footage to the instructions. Open the batch to inspect other episodes, their outcomes and quality flags.

DROID / Episode 14AI draft

Episode annotations

Interval (s)StepInstruction
0.00–5.07reachMove the arm toward the towel.
5.07–6.00graspGrasp the towel.
6.00–11.87moveCarry the towel to the bowl.
11.87–12.87placeRelease the towel inside the bowl.
12.87–14.47retreatRetreat the gripper upward.
Explore the batch ↗

What you get

Inside a labeled batch.

Every interval links a movement to its instruction and its place in the recording.

steps.jsonSample · episode 14
step
move
start_s
6.00
end_s
11.87
evidence
gripper closed, arm carrying

Try the original footage.

Switch cameras or jump to a labeled interval. The video and labels stay on the same clock.

reach5.07 s
grasp0.93 s
move5.87 s
place1.00 s
retreat1.60 s
DROID / Episode 14Front camera
AI draftreach
Move the arm toward the towel.
0.0 s

Inside the demo batch.

Inspect the latest counts from the public DROID demo. Select a breakdown to see what the batch contains.

Inspect the episodes ↗
DROID / Public batchSnapshot on load
—episodes
—minutes of footage

How many attempts completed the task?

Loading batch data…

Will your dataset fit?

Bring a representative task to the first conversation. A short description is enough to start.

01

Format

Annota works with LeRobot datasets. For HDF5, discuss the layout with us first.

02

Signals

Tell us which camera views, gripper signals and arm motion records accompany each episode.

03

Success criteria

What counts as a clean completion? Which recoveries should stay in training? These choices shape the review.

Try the demo with a question.

The public DROID batch contains AI drafts. Use it to see how the evidence and labels fit together.

towel to the bowlexample · episode 14

move · 6.00–11.87 s · “Carry the towel to the bowl.”

  1. 01Search for an instruction that interests you.
  2. 02Open an episode and inspect its video at a step boundary.
  3. 03Return to the report and see how often that step appears across the batch.
Open the demo ↗

Book a call with us.

Tell us what your team is training. On the call, we can discuss your dataset, the labels you need and whether Annota fits your project.

Our team reviews each request and follows up by email about scheduling.