Send an episode
Any ego camera, one short clip is enough. NDA on request; footage is used for your evaluation only.
The annotation API lands soon. Share one episode today; it comes back the same day, fully layered.
POST /videos/{id}/annotatepreview
{
"profile": "semantic",
"sampling": { "fps": 2 },
"hints": {
"task_description": "Focus on , , , ."
}
}Select a layer to update the video preview and delivery artifacts.
Hand pose 21×2source truth
3D keypoints projected via calibrated cameras
hand_pose.jsonlobjects.jsonldepth/steps.jsonoverlays/lerobot/
Every dimension, moving
Each cell plays one annotation layer on one scene, exactly as delivered. Retail, apparel, office, kitchen, workshop.
Depth is relative, not metric. The handloom trajectory is relative scale. Contact and grasp are review candidates, not tactile ground truth.
Try it before the API ships
The fastest way to judge annotation quality is on your own footage.
Get it back, layered
All applicable layers plus an overlay render per layer, in the same artifact structure the API will return.
Scale on your spec
Your taxonomy, not ours: step granularity, object classes and grasp categories defined per project.
samples@cerul.ai · attach a clip or share a link, we reply the same day
Lands in your training stack
One canonical layer set, exported to the formats your pipeline already reads. Ego human data ships with a documented pose-as-action mapping; retargeting to your embodiment stays on your side.
LeRobotepisode datasets for the HuggingFace ecosystem
openpicompatible out of the box via LeRobot format
RLDS / OXETFRecord episodes for Open X-Embodiment style pretraining
HDF5robomimic and ACT style containers
MCAPnative capture container, Foxglove ready
COCO / JSONLper-layer files for perception training
Why teams pick us
Data processing only
We do not run capture crews or sell model training. One step of the chain, done properly.
Verifiable, not claimed
Calibrated stereo capture means every layer can be re-checked against source truth. Provenance is printed on each layer.
Models do the heavy lifting
Self-trained annotation models handle throughput; humans review instead of drawing boxes. That is the price gap.
End-to-end when needed
Partner capture network. Bring a spec, get delivered episodes without managing collection yourself.
Questions teams ask
Who owns the annotated data?
You do. Annotation artifacts produced from your footage are delivered under your ownership; we keep nothing beyond what the engagement requires.
Can you work under NDA with private delivery?
Yes. NDA before any footage moves, private delivery channels, and your material never appears in our public samples without written clearance.
How fast is a sample turnaround?
Same day for a short episode across the standard layers. Larger batches are quoted with a delivery schedule up front.
Do we have to buy capture too?
No. Processing-only is the default; bring footage from any source. If you also need collection, we scope it through partner capture teams under one contract.
How is pricing structured?
Per annotated minute, priced by the layer set and review depth you choose. Sample evaluations are free; pilots are quoted against your spec.