What Is a Capture Session in Egocentric Data Collection?

A skill spec is the written contract for a capture program: the named skill, the environments and conditions to cover, the acceptance criteria and the target hours. The site states it is agreed before a single contributor is briefed and that nothing is captured speculatively. For a session, that means the task, the room and the pass criteria are fixed before anyone puts on a rig. Each episode is later checked against the named skill, and one that does not match is rejected as a spec miss.

Firsthand states hardware-triggered multi-sensor rigs on a single PTP clock domain, with head and wrist video, depth and IMU delivered as separate timestamped streams. Calibration is solved before and after the session, and the session does not start unless RMS reprojection lands under 0.28 px. Each session is supervised: a second operator verifies sync and framing between takes. Contributors sign a written release granting commercial training rights before capture. Session length, crew size and episodes per session are not stated.

The site has no standalone definition of the term, so this entry describes how the capture methodology page and the calibration guides use it. A session is the recording block in which an operator wears the rig and performs the named skill while a second operator supervises. Calibration, sync checks and supervision are bracketed around it.

For egocentric capture, the methodology page states that rooms are recorded as found and that the team does not build sets, because a staged kitchen teaches a policy that counters are empty.

An episode is one recorded demonstration of a skill, packaged with meta, streams, annotations and calibration. A session is the block of recording that produces demonstrations. The skill spec says what to capture, and a batch is a group of validated data delivered together. The site does not state whether a session holds one episode or many, so confirm that in your spec.

The checks sit at different levels. Calibration is solved at both ends of the session. Sync, occlusion, depth, consent and spec checks decide whether an episode ships.

Before: the spec fixes the skill, room and criteria, the operator has been trained, and contributors have signed releases. The opening calibration then runs: a target solve, per-camera intrinsics, extrinsics against the shared target expressed in the head-camera frame, and IMU alignment from a 60 s static hold plus a figure-eight. If RMS reprojection is not under 0.28 px, the session does not start.

During: every stream is hardware-triggered onto one PTP clock, and the second operator verifies sync and framing between takes.

After: the calibration procedure repeats as a closing solve. Rejection thresholds apply automatically at ingest, so no human decides a drifted session is probably fine. Failing episodes go to the reject log and are recaptured.

Passing episodes ship as validated batches. Each batch carries a data card, consent artefacts and a reject log, and coverage is reported against the skill spec as fill rates by condition. The data follows the canonical episode schema, projected into RLDS, LeRobot, WebDataset or HDF5. Billing is per validated hour: 60 minutes that passed QA against the spec, not raw recording.

The first validated batch arrives in a median of 48 hours from a signed spec where coverage is already deep, and in three to four weeks where contributors must be recruited first.

Fast, on-demand and custom collection are scoped by a written skill spec: the skill, environments and conditions, calibration fields and acceptance criteria. Session length, crew size, venue types and regions are not stated, so agree them in the spec. Request a custom dataset to start, optionally with a Spec Pilot.

What is a capture session?

One supervised recording block in which an operator wearing a calibrated rig performs a named skill in a real room. Its demonstrations are packaged as episodes and checked against the skill spec.

Is a capture session the same as an episode?

No. An episode is one recorded demonstration packaged with meta, streams, annotations and calibration. A session is the recording block that produces them. How many episodes one session yields is not stated.

How long is a capture session?

Not stated. The site gives no session length, crew size or throughput, so agree them in your skill spec.

What is checked at the start and end of a session?

Calibration is solved at both ends. The session does not start unless RMS reprojection is under 0.28 px, and drift between the opening and closing solve is written into episode metadata.

Is consent handled per session?

Contributors sign a written release granting commercial training rights before capture. An episode with a bystander who cannot be de-identified is rejected as a consent gap.

What happens when an episode fails a check?

It is rejected, logged in the reject log and recaptured.

Can I request custom collection sessions to my own spec?

Yes. Fast, on-demand and custom collection are scoped by a written skill spec, so request a custom dataset and list the skill, environments, calibration fields and acceptance criteria.