What Is Custom Data Collection?

A public or off-the-shelf dataset was captured to someone else’s spec, so it rarely matches a specific model’s exact failure mode, embodiment, or population. Custom data collection exists because the buyer, not the dataset publisher, defines what gets captured: which tasks, which environments, which objects, and which edge and failure cases matter for their own model. That control over the spec — and the ability to validate it on a small pilot before committing to volume — is what separates custom collection from picking a dataset off a catalog.

Firsthand runs custom data collection as a written-spec, on-demand service: a described gap becomes a spec with accept and reject criteria, a Spec Pilot proves the spec on real footage before volume scales, and the full program runs through contributor recruiting, capture, per-batch QA, consent documentation, and anonymization. Delivery arrives under a perpetual, buyer-owned commercial license in RLDS, LeRobot format, WebDataset, or another format the training pipeline already expects.

A custom data collection project starts with a described gap — a task, environment, object, or population a current model is missing or failing on. That gap becomes a written skill spec covering the tasks, environments, objects, coverage, and accept and reject criteria a batch has to clear, so both sides agree on what "done" looks like before any contributor is recruited.

Most programs then run a Spec Pilot: a small, fast first batch captured against that exact spec, used to confirm the spec and QA bar hold up on real footage before the program scales to full volume. From there, the collection moves through contributor recruiting and consent, capture, per-batch QA against the written accept criteria, anonymization of faces, plates, and screens, and delivery of the finished, licensed dataset in the buyer’s training format.

A public or off-the-shelf dataset is a fixed artifact, already captured to cover whatever task distribution its original publisher or buyer needed. It can be a fast way to start, but it was not built around any specific model’s current failure mode, and its license terms are set by the publisher rather than negotiated by the buyer.

Custom data collection instead starts from the buyer’s own spec: the tasks, objects, and environments captured are exactly the ones the buyer names, scoped around the coverage gaps a public corpus does not already fill. The tradeoff is time to first data — a custom program still has to run a pilot and scale capture — against a dataset that already exists but may not match what the model actually needs.

A data labeling service takes footage or sensor data a buyer already has and adds annotations — bounding boxes, segmentation, action labels — to existing recordings. Custom data collection is upstream of that: it captures new footage in the first place, against a written spec, rather than labeling data that already exists.

The two are often used together inside one engagement — a custom collection program includes its own annotation step as part of delivering a finished, trainable dataset — but a buyer that already has raw footage and only needs it labeled is better served by a labeling service than by a new collection program.

Custom-collected data ships under a perpetual, buyer-owned commercial license on delivery, documented through a signed consent chain from every contributor. That is different from many off-the-shelf datasets, whose publisher license terms are often non-exclusive and set by the publisher rather than negotiated by the buyer.

Timelines depend on how deep existing contributor coverage already is for the spec’s tasks, environments, and regions. Where coverage is already deep, a signed spec can reach a validated first batch in a median of 48 hours. Where coverage is thin and contributors need to be recruited first, a realistic timeline runs three to four weeks. Either way, a Spec Pilot — a small first batch — is the normal starting point before a program scales to full volume.

What is custom data collection?

Custom data collection is commissioning new training data captured to a buyer’s own written skill spec, rather than selecting from a fixed, already-captured dataset. A pilot batch validates the spec before the program scales, and delivery ships under a buyer-owned commercial license.

How does custom data collection differ from data collection as a service?

Custom data collection describes the activity — capturing new data to a spec. Data collection as a service describes the delivery model: a managed provider running that entire activity end to end on the buyer’s behalf, including spec, contributors, capture, QA, and delivery.

How does custom data collection differ from on-demand data collection?

Custom data collection is the broader concept of capturing data to a spec. On-demand data collection adds a speed dimension: a provider’s already-recruited contributors and warm pipeline mean a new brief does not start from zero, so a custom program can move from spec to first validated batch quickly where coverage is already deep.

Is custom data collection the same as egocentric data collection?

No. Egocentric data collection names the capture modality — first-person video from a camera worn by the person doing the task. Custom data collection describes the commissioning model — capturing new data, of any modality, to a buyer’s own spec — and egocentric capture is one modality that can be custom-collected.

Do I need a finished spec before starting a custom collection?

No. Describing the gap — a task, environment, or population the model fails on — is usually enough to start. A written spec with accept and reject criteria is one of the first things the engagement produces, not a prerequisite for starting one.

What happens if footage doesn’t meet the spec?

Every batch is reviewed against the spec’s accept and reject criteria. Footage that doesn’t clear the bar is rejected and recaptured, rather than delivered with a caveat.

Can a custom data collection be topped up later?

Yes. A top-up runs against the original spec and accept criteria, so new data stays consistent with what a buyer already holds and slots into the same schema, without renegotiating the spec from zero.

What license does custom-collected data come with?

A perpetual, buyer-owned commercial license on delivery, documented through a signed consent chain from every contributor — different from many public or off-the-shelf datasets, whose license terms are set by the publisher and are often non-exclusive.