subtask · 2.200–4.800 s

Repositioning left hand and adjusting warp threads across the reed

event · pull · thread → loom

hands · 21 keypoints · frame 118 / 210

OPEN SOURCE · NO ACCOUNT · YOUR MODELS

Turn experience captured on video into machine capability.

0 sepisode 6774a70f/0 · 3.900 s7.0 s

subtask

event

interaction

state

TASKSUBTASKEVENTINTERACTIONSTATEFLAGPROGRESSHANDS TASKSUBTASKEVENTINTERACTIONSTATEFLAGPROGRESSHANDS

timed labels from a 7-second clip, one command.

210

hand frames tracked on CPU, no key, nothing uploaded.

5/5

episodes pass the official LeRobot loader after writeback.

accounts. Download one binary and start.

One episode, read like a score.

Every label sits on the episode’s own clock, in integer microseconds, so it lines up with frames, actions and state.

Separating warp threads across the reed

Repositioning left hand, adjusting threads

Retract

Reaching forward, adjusting threads

0.0002.2004.8005.4007.000 s

Observe

Read LeRobot v3.0/v3.1 episodes or any video. Track human hands locally.

annotate ./dataset --hands

Understand

Your vision model writes subtasks, events, contacts and state changes.

annotate ./dataset --semantic

III

Write back

Review it as video, then put subtasks into a new dataset copy for training.

--write-lerobot --out ./labeled

From the showcase

See every episode →

LOOM · 7.0 s · 11 labels · 210 hand frames

Hand weaving, first-person

event · pull · 5.9–7.0 s

Reaching forward

No sign-up.
No hosted API yet.
Just a binary.

macOS Apple Silicon and Linux x86_64. Bring a Gemini key or any compatible endpoint. Apache-2.0.

$ curl -fsSL https://github.com/cerul-ai/cerul-robotics/releases/latest/download/cerul-robotics-installer.sh | sh

$ npx skills add cerul-ai/cerul-robotics # for coding agents

Every demonstration, readable.

Install Cerul Robotics Need labels delivered? Talk to us