Egocentric Napkin and Towel Folding Data

Egocentric napkin and towel folding data is real-world video of a person folding table napkins and folding towels and linens, recorded from the operator point of view while they work. The footage shows the hands, the cloth and the surface under it. Alongside the video, each episode follows the Firsthand episode schema, with depth, 3D hand pose, contact events and frame-accurate action segments aligned on one hardware clock.

Teams search for it under several names, including napkin folding robot training data, egocentric towel and linen folding video, a hospitality folding dataset and custom hospitality data collection. They all describe demonstrations of the same two folding jobs, collected to a written spec. This page covers only what the Firsthand site states about them, and marks everything else as Not stated.

Two rows in the site task catalog cover this work, both in the Hospitality Service Set environment under Hospitality & Events. The first is "Fold napkins", described as "Fold napkins to a set style." and rated Medium. The second is "Fold towels + linens", described as "Fold towels and linens; stack by type." and rated Medium. Both use a multiview rig with handheld grippers, and both are listed as Available on request. The rows are defined in lib/commercial-catalog-data.ts.

Anything outside those two rows is Not stated. That includes the names or number of napkin fold styles, the sizes, fabrics or colors of the cloth, the types a stack is sorted into, and any named brand, linen supplier, hotel or restaurant, so none is claimed here. If your policy needs one of them, you write it into the spec and we scope it, because the catalog rows are a starting vocabulary, not a fixed menu.

The site defines Available on request as work that is scoped and recruited to spec when a buyer asks for it. Neither row is listed as Active collection or Ready to deploy, so the site does not say that a napkin or towel folding capture stream is running today, and it does not state delivered hours for either row. A first batch is quoted against your spec rather than drawn from a stream that already exists.

The site does not list an off-the-shelf napkin, towel or linen folding dataset, so this is custom collection. Custom collection at scale runs through a network of thousands of capture providers, for a consistent, recurring data flow to labs.

The catalog names the environment Hospitality Service Set (table, bar, room) and gives its examples as a table, a counter or bar, and a housekeeping or room area. It is a staged working set for hospitality service tasks, not a described real venue. The same environment lists twelve rows in total, among them Set table service, Clear + bus table, Tray assembly + delivery, Wipe + reset counter and Make up room bedding.

The site names no hotel, restaurant, linen service or building, and does not state the number of sites, the regions or the venue type. It does not state capture at a customer own facility for these rows either. Environment, access and schedule are agreed in the written spec.

The catalog states one capture mode for both rows: a multiview rig with handheld grippers. The site describes this mode as a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers, so demonstrations map onto a robot end effector. It also states that every sensor in a multiview rig, meaning the head camera, both wrist cameras, depth and the IMU, is hardware-triggered from a single PTP grandmaster clock.

The site does not state a head-mounted capture mode for these two rows, so this page does not claim one, and it does not claim that a stereo pair is running on them. Stereo capture is a general Firsthand offer that is agreed in the skill spec. A per-environment stream list is Not stated, so which of the ten streams an order carries is decided in the spec. Camera count, resolution, frame rate, table layout and cloth volume are Not stated, so raise any of them in the spec.

Several pages on this site touch folding or linen without being about these two rows. The household laundry folding page covers folding and sorting in homes. The commercial laundry and linen page covers sorting, loading, folding and pressing across a chain of laundry stations. The table setting and clearing page names Fold napkins only as a neighboring row. The hotel room turnover page covers resetting a guest room, and it treats laundry rows such as folding as back-of-house laundry rather than the room itself.

This page narrows to the two Hospitality Service Set rows that name folding napkins and folding towels and linens, and it describes the capture mode those rows use. The comparison table below sets each neighbor against this page. A buyer who needs the whole laundry chain starts from the commercial laundry page. A buyer whose policy ends at a folded napkin or a stack of folded towels on a hospitality service set starts here.

Each episode follows the Firsthand episode schema: ten streams synced under 2 ms, with head and wrist RGB, depth, 3D hand pose with per-joint visibility, contact and grasp events, and frame-accurate action segments. A fold is a run of contact events with a hand often covering part of the cloth, and the site pairs that with 3D hand pose that carries per-joint visibility, contact events on grasp and release, an IMU on one clock and action segments that mark where one fold ends and the next begins.

Object labels use the terms your spec sets. Napkin styles, cloth types, stack categories and the number of folds per item are Not stated. Task durations, cycle times and the number of episodes are Not stated too. Success or quality labels, such as whether a napkin matches a style or a stack is sorted correctly, are Not stated, as are certifications and robot results. The site makes no safety-compliance claim for this work.

The catalog rates both rows Medium, so the site does not call either one Hard, and this page makes no claim about robot results. What the site does say is where cloth work is difficult to capture. On its laundry folding page it describes fabric that changes shape with every touch, folds over itself and is half hidden by the hands doing the folding. On its commercial laundry page it adds that fabric covers fingers, so per-frame instance masks and a per-joint visibility flag on hand pose mark what is hidden instead of guessing it.

The viewpoint matters too. The site states that a robot or humanoid perceives from a head- or wrist-mounted camera, not a static overhead shot, and that egocentric capture matches that geometry. The two rows here are recorded by a multiview rig that includes a head camera and both wrist cameras, so the geometry a buyer receives is the one the rig and the spec define.

A good spec treats each named job as its own task. Folding a napkin to a set style gets a start state, an end state and accept criteria, and the spec names the style because the site does not. Folding a towel or linen and stacking it by type gets its own start and end states and says what counts as a type and what counts as a finished stack. Naming the end states keeps action segments consistent across contributors and lets you ask for more of one job without re-collecting the other.

The same environment lists Set table service, which lays a full place setting, and Make up room bedding, which changes a case and spreads a cover. A spec can chain a folding row with them into one hospitality sequence if your policy needs the whole run. Whether such chained sequences are collected as one episode or several is agreed in the spec.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. These episodes use the same schema as every other Firsthand collection, so a loader that reads another Firsthand dataset reads this one without changes. The format table below sets each format against what its official documentation states.

Every batch is validated against signed accept criteria before it ships, with a reject log, and delivery is rolling, per validated batch. Rejected items are recaptured at our cost, and billing is per validated hour.

Every episode links to a signed contributor release through a documented consent chain, and the consent record is part of the delivery. A vision model runs on every frame at ingest, flagging faces, heads, licence plates and any legible screen or document. Those regions are irreversibly blurred, and a human reviewer confirms the result before footage ships. If a clip cannot be reliably de-identified, it is discarded rather than shipped, and the anonymization record is delivered with the data.

Anyone else in frame, such as a guest, a co-worker or a bystander, is handled through the site bystander and de-identification process rather than being captured as a second consenting participant. The site also states that front-of-house hospitality settings are bystander-dense, which lowers yield after de-identification, while it classes room work as back-of-house. Which setting a folding job is recorded in is agreed in the spec. The site states no other privacy, compliance or legal terms for this work, and no in-home capture for these rows.

Where coverage is already deep, the first validated batch lands in a median of 48 hours from signed spec. The site classes hospitality as a thinner-coverage environment, where vetting operators for guest-facing and back-of-house roles is the bottleneck rather than demand for the task. A hospitality batch therefore typically takes three to four weeks, and we quote that up front. Both rows are Available on request, so plan on a quote rather than assuming the 48 hour figure.

The site also states program milestones. A pilot reaches its first delivery in two to three weeks, and a standing program reaches full rate in four to six weeks. Those assume capture coverage is already in place. On-demand top-ups run against the same spec and accept criteria after each training run. A batch size and a later-batch turnaround are Not stated, so both are quoted against your spec.

You describe the folding jobs, the setting and the sequence your policy needs. We turn that into a written spec covering tasks, environments, capture tier, annotation depth and accept criteria, then quote reach, timeline and price before any capture begins. This is the custom egocentric data collection route that every Firsthand task family follows.

A Spec Pilot, a small and fast first batch, can prove the spec and accept criteria before you scale. Every collection ships under a perpetual, buyer-owned commercial license, with a reject log and signed accept criteria on every batch.