01
Collect
Calibrated head rigs, wrist IMUs and optional depth. Recorders work from a task script with a target count of episodes and a required variation list.
Platform
Collect. Parse. Filter. Deliver.
Raw first-person footage is close to useless — shaky, over-exposed, full of bystanders and repeated takes. The value is in what happens after the record button. Here is every stage, and what you get out of each.
02
Parse
Sync, de-warp, stabilise. Then hand pose, gaze projection, object tracks, contact events and a narration line per action segment.
03
Filter
Automatic rejection for motion blur, tracking loss, exposure blowout and near-duplicate takes; manual review for consent and framing.
04
Deliver
Packed episodes plus a data sheet: provenance, consent references, filter statistics and the gaps we know about.
Capture stack
What a rig records, per second.
- Video
- 1440p at 60fps, 120° horizontal field of view, rolling-shutter corrected. Stereo option on select rigs.
- Gaze
- Binocular eye tracking at 120Hz, projected into image space with per-frame confidence.
- Hands
- 21-point pose per hand, left/right identity, plus contact and release events against tracked objects.
- Motion
- Head IMU at 200Hz; wrist IMU at 100Hz; 6-DoF head pose from visual-inertial odometry.
- Audio
- Two-channel spatial audio, with speech muted by default unless the task requires it and consent covers it.
- Depth
- Optional time-of-flight depth at 30Hz for grasp-critical collections.
Filtering
We throw away two-thirds of what we shoot. That's the product.
A model does not improve because you fed it more hours. It improves because the hours were different from each other and clean enough to learn from.
Our filter stage scores every episode on tracking integrity, exposure, occlusion balance and novelty against everything already in the collection. Episodes that only repeat what the set already contains get dropped, no matter how well shot they are.
Filter report
- Submitted: 612 episodes / 41.3 hrs
- Rejected — tracking loss: 88
- Rejected — blur or exposure: 61
- Rejected — consent or framing: 24
- Rejected — near-duplicate: 251
- Released: 188 episodes / 12.6 hrs
Keep rate 30.7% · typical range 26–36%
Delivery
It arrives in the format your training run already reads.
Schemas
LeRobot, RLDS, WebDataset shards, or a custom schema you define. Annotations as JSON or Parquet.
Transport
Your S3 or GCS bucket, an SFTP endpoint, or encrypted physical drives for large one-time transfers.
Data sheet
Per-collection documentation modelled on Datasheets for Datasets: how it was made, who by, what it under-represents.