Egocentric Duvet Cover Insertion Data

It is real-world video of a person putting a duvet into its cover, with synchronized sensor data alongside, captured to a written skill spec. The buyer sets the task and the accept criteria; Firsthand handles contributors, capture, QA, consent and delivery. This page is about egocentric duvet cover data in the sense of what a policy can learn from watching a person do this step, and it makes no claim that any robot has been trained on it.

The data is human capture by paid, consented contributors, not robot or teleoperation data, and it is not a fixed corpus you download. The catalog row is "Insert duvet cover", described as "Insert a duvet into its cover." The subject of this page is that one step. The wider job of making a bed, stripping linen or changing a case has its own pages, compared in the table below.

One. In the Deformables & Textiles Bench environment of the Manufacturing & Workshops sector the catalog lists "Insert duvet cover": "Insert a duvet into its cover." It is rated Hard, its status is Available on request, and the capture setup it lists is Multiview rig + handheld grippers. The environment examples read "Cloth-and-cable bench: garments, towels, cables, form boards, bags." The first table below reproduces the row exactly.

Available on request means the work is scoped and recruited to spec when a buyer asks for it. This page therefore claims no completed collection, no delivered hours and no delivered footage for this row. Four other rows are context only: "Change bed linens" in the Care Room, "Make up room bedding" in the Hospitality Service Set, "Make a bed" in the Care Room and "Hotel Bed Making" in Hotel & Lodging. They are listed in the same table and are not the subject of this page.

The site describes two capture modes for its catalog rows. In the first, a head-mounted phone records the first-person view and a multiview rig adds synchronized third-person angles when the spec asks for them. In the second, a staged working set is recorded by a multiview rig while the operator performs the task with handheld grippers, so demonstrations map onto a robot end effector. "Insert duvet cover" is listed with the second: Multiview rig + handheld grippers.

Read plainly, that means this row is a staged bench task, not a recording of someone changing their own bedding at home. How the first-person viewpoint is arranged inside that setup, which cameras are used and whether a head view is part of it are Not stated on the row and are agreed in the skill spec and confirmed on a Spec Pilot. The site states no stereo pair for this row, so this page describes egocentric and multiview capture only.

Three problems are already described on this site for deformable work. The first is hand occlusion. Fabric covers fingers, and the site states that 3D hand pose carries a per-joint visibility flag so a hidden joint is flagged as hidden rather than silently guessed, with contact events marking when a hand meets an item. A duvet going into a cover hides what is inside it as well as what is holding it, so the labels have to say what is visible rather than imply a clean view.

The second is the mask. Masks propagated from frame to frame work for rigid objects and not for cloth, which is why per-frame instance masks, stored at the depth resolution of 848×480 so they line up with the depth stream, are shipped where the spec calls for them. The third is the length of the step chain: the catalog lists this as a Hard row, and the site does not state why, so this page does not guess. Lighting for the bench is Not stated and is fixed in the spec.

Depth is also worth naming. The site states that depth degrades on wet and reflective surfaces and that per-frame confidence maps mask low-confidence regions. Whether a given duvet or cover fabric behaves differently is Not stated, so the Spec Pilot checks the real footage before the order scales.

Each attempt is delivered as one episode, and a duvet order needs a decision up front about what one attempt is. The row description gives a start and an end: a duvet, and the same duvet inside its cover. Whether an episode begins with the cover laid out, whether it includes straightening the result and whether several attempts share one recording are spec decisions, because the site does not state an episode length.

Inside the episode, the synchronized streams run from start to end and action segments divide it into labeled spans, each with start_ns, end_ns, verb and noun. A spec for this row names the verb and the noun rather than a broad activity, as the site’s skill-spec guide advises. The verbs and nouns are written into the spec for your program; this page does not supply a step list or any instruction on how the task is done.

For episodes in general the site states ten hardware-synchronized streams, synced under 2 ms, including head and wrist cameras, depth with confidence maps, IMU, 3D hand and body pose and instance segmentation. Hand pose is 21 joints per hand with a per-joint visibility flag. Calibration is measured every session and shipped with the episode, with a 0.28 px RMS reprojection gate. The language data is one task instruction per episode.

Which of these a "Multiview rig + handheld grippers" capture carries is Not stated for this row. In particular, whether a gripper capture delivers hand pose in the same form as a head-mounted capture is not stated, so it is asked for in the spec. The fourth table below lays out what the site states and what it does not.

This page is about one step on one catalog row. The egocentric bed making page covers the bed as a whole: stripping it, fitting and tucking sheets, changing a case or a cover, placing a duvet and arranging pillows, and it lists "Insert duvet cover" as one of its available-on-request rows. That page is the one to read for the full bed job; this one is for a spec that wants the duvet step on its own.

The laundry folding page centres on folding clean laundry, and the commercial laundry and linen page covers linen handling in a commercial laundry and hotel linen service. The stereo hotel room turnover page covers room reset, bed making and vacuuming as stereo capture, with four rows all listed as Ready to deploy. The egocentric deformable object page covers soft items in general. The second table below puts them side by side.

Every batch is validated against signed accept criteria before it ships, with a reject log. The site states that an episode is rejected when the manipulated object is out of frame for more than 15 percent of it, when more than 35 percent of depth pixels in the manipulation region are invalid, or when an undocumented bystander appears. Episodes that fail are rejected and logged per batch, not delivered with a caveat.

Billing is per validated hour, and rejected items are recaptured at our cost. For a large item that spends part of the attempt inside another, the out-of-frame rule is worth reading against your spec, because what counts as the manipulated object is a decision the spec settles. The site states no accuracy figure for step labels, so what counts as a pass for a step boundary is written into the accept criteria.

The site’s off-the-shelf catalog lists about 324,800 hours across 18 datasets. None of its descriptions names duvet cover insertion. One stereo set, S2 (~5,000 hours, Cambodia, 2, 4 and 6 camera rigs, stereo RGB, task labels and hand keypoints), lists bed making among its tasks. Whether S2 holds any duvet step is Not stated, so it is a way to look at a format and rig, not coverage of this row.

Duvet cover collection to a spec is therefore custom collection: custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. Where coverage is deep, the median time from signed spec to first validated batch is a median of 48 hours; where it is thin, turnaround runs three to four weeks. The row is Available on request, so the scoped-to-spec timeline applies rather than the ready-now case, and the quote states it up front. A published model write-up, GEN-0, linked in the sources, says "data quality and diversity matters more than sheer volume", which for this row points at the spread your spec covers rather than a raw count.

Every episode links to a signed contributor release through a documented consent chain, and the consent record is part of the delivery. Faces, plates and screens are detected and irreversibly blurred at ingest, and a reviewer confirms the result before anything ships, with the anonymization record delivered alongside. Speech is scrubbed of names, addresses and card numbers, and footage that cannot be de-identified is discarded.

The buyer owns the dataset under a perpetual, irrevocable, worldwide, non-exclusive-by-default licence, with exclusivity available as a separate term. Capture at your own facility is listed as not offered, so the setting is one Firsthand arranges against your spec.

Duvet size, weight and fill, cover material and closure type, bed size, mattress type, fabric or product type, brand, the number of items per episode, task times, success rates, error rates, force or tactile data, a cloth-state or wrinkle label and a task-completion score are not stated on the site as standard fields. No customer, property or facility is named for the row.

If your policy needs any of these, they are written into the skill spec as annotation depth and covered in the quote before capture begins. This page gives no instructions for how the task is performed.

Name the row you are following, "Insert duvet cover", the start and end state of one attempt, what counts as a reject, and the capture mode: the listed multiview rig with handheld grippers, or another arrangement you want checked on a Spec Pilot. Then say how many variations of duvet, cover and setting your policy should see, which streams and labels you want, and which delivery format your pipeline reads.

Delivery can be RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or raw streams with metadata, and MCAP is also offered. Most programs begin with a Spec Pilot, a small first batch that proves the spec and accept criteria on real footage before you scale. The last table below shows what each spec element decides.