Order a Custom Egocentric Dataset

Ordering starts with a written spec: the tasks or skills to capture, the scenes and environments, the objects and contributor demographics, the streams and labels you need, and the accept criteria a take has to clear to count as validated. That spec is what gets quoted — not a catalog price — because reach, stream count, and turnaround all move the number independently.

Most orders start with a Spec Pilot: a small, fast first batch that proves the written spec and accept criteria actually hold on real footage before you commit to volume. It surfaces whether the reject rate is manageable, whether coverage for your tasks and regions is thin or deep, and whether the stream count you asked for is actually necessary.

Once the pilot clears, the same spec scales to the full order. Capture, review, and delivery run in batches rather than one large drop, so you can see validated data arriving and QA passing before the program is fully delivered.

A usable spec answers: which tasks or skills, in which environments and regions, performed by which kind of contributors, captured with which streams (video, depth, IMU, hand pose, body pose, segmentation, action labels), and checked against which accept criteria. Coverage and volume are defined by you in the spec — how many hours, how many contributors, how many environment variants — rather than sold as a fixed package size.

The spec also needs a delivery format: RLDS, LeRobot, WebDataset, or whichever structure your training loader already expects. Specifying it up front means the dataset arrives ready to load rather than needing a conversion step after delivery.

A delivered order includes the synchronized streams and labels defined in your spec, a QA report showing what was reviewed against the written accept criteria, documentation of the consent chain for the footage included, and the dataset itself in the agreed delivery format. Faces, license plates, and screens are detected, blurred, and reviewer-confirmed as part of QA before anything ships.

Everything that is delivered has already cleared the accept criteria in your spec — raw takes that fail review are rejected and recaptured before delivery, not shipped and flagged afterward.

Where coverage for your tasks, environments, and regions is already deep, the median time from signed spec to first validated batch is 48 hours. Where coverage is thin — a region, task, or condition combination with little standing contributor coverage — turnaround runs three to four weeks, because recruiting has to happen in parallel across more regions before capture can scale.

A Spec Pilot gives you a real, small-scale timeline number specific to your spec before you commit to the full order, rather than relying on a general estimate.

There is no fixed price list. Cost is driven by how many synchronized streams each episode needs, how many validated hours the order requires, how thin or deep existing coverage already is for your tasks and environments, and how fast you need the first batch. Billing is strictly per validated hour: footage that fails QA is rejected and recaptured at the provider’s cost, not billed to you.

See the cost-driver breakdown on the pricing page for the full explanation of how each of these levers moves the number, and why a Spec Pilot is the fastest way to get a real figure against your spec rather than a rough estimate.

A public egocentric dataset is fixed: whatever tasks, environments, and contributor mix it already contains is what you get, under whatever license its publishers attached, which is frequently research-only and not cleared for commercial model training. A custom order is scoped to the exact gap your model has — the tasks, objects, regions, or contributor demographics missing from what you already hold — and ships under a license written for commercial use.

Buying custom makes sense when the gap is specific enough that no public corpus covers it, when you need a commercial license a research-only dataset does not grant, or when you need the data on a timeline a static public release cannot move to meet. If a public dataset already covers the gap under a license you can use, there is no need to commission a custom order for it.

Delivered data ships under a perpetual, buyer-owned commercial license. The buyer is not renting access to a hosted corpus or limited to research use — the license is written to permit training and deploying commercial models on the delivered data.

Delivery ships in RLDS, LeRobot, WebDataset, or other formats your training loader already expects, specified as part of the order rather than fixed to one structure. Matching the format to your pipeline up front avoids a conversion pass between delivery and training.