Frontier AI labs
Requirements
Frontier AI training data has to target the behaviors a model is actually missing. A custom spec fixes tasks, environments, and regions to your eval, so custom egocentric data for frontier labs lands where the loss is.
Ten synchronized streams per episode (4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation, and action labels), hardware-aligned under 2 ms.
Every item is measured against signed accept criteria and reviewed before delivery. You are billed per validated hour, and each batch ships with its reject log.
A perpetual, buyer-owned commercial license backed by a documented consent chain, so a model trained on the data can go into a product.
Off the shelf
An off-the-shelf dataset was scoped for someone else's model. Its tasks, environments, and regions were fixed before your failure modes were known, and every buyer trains on the same hours, so it rarely closes the specific gap a frontier model is failing on.
Rights are the second gap. Before a commercial training run you have to confirm that each dataset permits commercial use and that its consent covers it. A custom collection is licensed to you and backed by a documented consent chain from day one.
The spec
You describe the behavior your model is missing. We turn it into a written spec covering tasks, environments, regions, capture tier, annotation depth, and accept criteria, then quote reach, timeline, and price against that document before any capture begins.
The collection runs through the same seven-stage end-to-end custom collection pipeline as every Firsthand program: contributor recruiting, capture, per-batch QA, anonymization, and licensed delivery. Rejected items are recaptured at our cost.
Fast · on demand
Where coverage is already deep, the first validated batch lands in a median of 48 hours from signed spec. Thin environments such as construction, clinical, agriculture, and hospitality typically take three to four weeks, because vetting operators there is the bottleneck, and we quote that timeline up front.
After the first batch, delivery is rolling. On-demand top-ups run against the same spec and accept criteria after each training run, so new data stays consistent with what you already hold.
Rights
Every custom collection ships under a perpetual, buyer-owned commercial license, and exclusivity is available so the data is never resold. Each episode links to a signed contributor release through a documented consent chain.
Faces, plates, and screens are detected and irreversibly blurred at ingest, and a reviewer confirms the result before anything ships. The anonymization record is delivered with the data.
Requirement vs delivery
| Requirement | What Firsthand delivers |
|---|---|
| Streams | Ten synchronized streams: 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation, action labels |
| Sync | Hardware-synchronized on one clock, under 2 ms across streams |
| Annotations | 3D hand and body pose, instance segmentation, action segments, grasp and contact events |
| Quality | Measured against signed accept criteria; billed per validated hour; reject log with every batch |
| Timeline | First validated batch in a median of 48 hours (deep coverage); three to four weeks in thin environments |
| Consent | Signed contributor release per episode; documented consent chain; faces, plates, and screens blurred and human-reviewed |
| Licensing | Perpetual, buyer-owned commercial license; exclusivity available |
| Formats | RLDS, LeRobot, WebDataset, plus HDF5, zarr, and Rerun .rrd |
Every collection runs through the same seven-stage end-to-end custom collection pipeline.
FAQ
Sources
Tell us what your model is missing. We will quote reach, timeline, and price against a written spec, with a first validated batch in a median of 48 hours where coverage is deep.