Egocentric Video Data of Office and Desk Tasks

Egocentric video data of office and desk tasks is first-person video of people doing everyday office work: working at a desk with documents, a keyboard and a phone, filing, handling mail, printing, scanning, operating a coffee machine, or running a desk-and-document station with a stapler, a binder and a label maker. Paid, consented contributors wear a capture rig and perform the tasks named in a written skill spec.

It sits next to two pages that already exist on this site without duplicating either. The workplace task page covers warehouse, retail, automotive, kitchen and assembly settings. The enterprise workflow page turns a hands-on SOP into labeled step units, and it describes a purely cognitive or paperwork process with no physical manipulation as a poor fit. Office and desk tasks are the physical side of paperwork: the hands feeding a scanner, squaring a stack, sealing an envelope, signing and filing a form. That hand work is exactly what a first-person camera sees well.

A robot or assistant that will work in an office needs to see office tasks the way the person doing them sees them. Footage from across the room shows that someone is working; it does not show how the hands load a paper tray, clear a jam, thread a form into a folder or seal an envelope. A head-mounted camera records the reach, the grasp and the hand-to-item relationship from the viewpoint a mobile manipulator or humanoid would have.

Office tasks are also a gentle entry point for manipulation research. Every task in the stated office catalog is rated Easy or Medium, the items are light and rigid or thin, and the setting is controlled. That makes them useful for teams who want varied, repeatable, short-horizon tasks, and for multimodal teams who want action-labeled video of ordinary indoor work.

Because the collection is custom and on demand, coverage follows your task list rather than whatever happened to be recorded. A first validated batch arrives in a median of 48 hours from signed spec where coverage is deep, and in three to four weeks where contributors have to be recruited into an unfamiliar environment.

Paper is thin and flexible, so a sheet or a stack changes shape every time it is touched, lies against other sheets and hides what is under it. Small items such as staples, labels, tape and envelope flaps are easy to lose behind a fingertip. The hands themselves occlude much of what they hold, and tasks such as collating or squaring a stack repeat many near-identical motions that only differ in order.

Typing and mouse work are small, fast finger movements that are hard to read even for a person watching, and the result of the work lives on a screen rather than in the hands. A real office also puts identifying material into frame by accident: names on monitors, forms on a desk, badges, mail and colleagues passing behind the contributor.

The capture program answers these with the streams and labels already stated on this site rather than with new ones: a calibrated head-mounted stereo view with depth, two wrist cameras, 3D hand pose with per-joint visibility, contact events on grasp and release, ordered action segments, and an anonymization pass for everything else in frame. The table below maps each challenge to what is stated, and says so where nothing is.

The site lists three environments in its Office and IT sector. Office and Professional Services covers six tasks: coffee machine operation, desk work, filing, mail handling, printing and scanning. Office Document Station covers twelve: assemble a binder, hole-punch and bind, label-maker and apply, load a tray and clear a jam, pack and tape a mailer, refill a stapler and tape, shred documents, sign and file a form, sort into folders, staple and collate sets, stuff and seal envelopes, and trim with a guillotine.

The third environment, Data Center and IT Hardware Bench, belongs to the same sector but is hardware work rather than paperwork: applying and scanning an asset tag, patching cables, seating a transceiver, reseating a card, mounting rails. It has its own page, linked below. All three are listed as available on request, and the office task rows are rated Easy except three document-station tasks rated Medium: refilling a stapler and tape, signing and filing a form, and stuffing and sealing envelopes.

Desk work is described in the catalog as working at a desk handling documents, a keyboard and a phone through routine office tasks. That is the task closest to what people mean by computer-use video, and a later section is explicit about what it does and does not include.

The same stream set as any other custom egocentric collection: ten synchronized streams including a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation and action labels. Depth is recovered from the calibrated stereo rig, the streams are hardware-triggered onto one PTP clock domain and aligned under 2 ms, the IMU is 6-DoF at 200 Hz, and hand pose is 21 joints per hand in 3D with a per-joint visibility flag.

The catalog lists the capture configuration per environment: Office and Professional Services is captured with a head-mounted phone or a multiview rig, and Office Document Station with a multiview rig plus handheld grippers. Which of the ten streams a given order carries is decided in the skill spec; the catalog does not state a per-environment stream list, so that is marked as not stated in the tables below.

Every episode also ships with its record: the consent chain for the contributor, the anonymization record, and provenance and data-processing documentation for your legal team. Calibration runs twice per session and is accepted only under the 0.28 px RMS reprojection gate.

Write the task as a skill spec: the tasks to demonstrate, the workstation and its layout, the objects involved, which streams and labels the pipeline needs, the accept criteria, and the delivery format. A desk-and-document task is specified the same way as any other, and only the setting, tasks and objects change.

Task decomposition does the real work. The catalog describes sign and file a form as writing, signing and filing a form, which a spec splits into ordered sub-steps. Each becomes a labeled action segment with a start and end timestamp, a verb and a noun, so a continuous run still yields individually labeled steps. Name the items too: tray, binder, mailer, stapler, shredder, label maker, guillotine, folder, cabinet.

Also decide what variation you need: different desks and rooms, different office types, different people, different starting states such as a jammed tray, a stack out of order or an unsealed envelope. Where a task is prone to going wrong, ask for failure and recovery cases, which the spec can require deliberately. The spec is where you settle which streams you need, how many contributors and how much volume, and what counts as a correct completion.

The anonymization page states that a vision model runs on every frame at ingest, flagging faces, heads, licence plates and any legible screen or document. Faces, plates and screens are then irreversibly blurred, and a human reviewer confirms the result before footage is delivered. The documents class is described as phone and monitor content, forms, badges and mail that would leak names, addresses or account data. That is the class an office puts in frame.

People are handled alongside it. Every contributor signs a documented release as part of a consent chain, the record travels with the data, and an anonymization record is delivered alongside it. Bystanders such as colleagues passing behind the desk are blurred at ingest, and clips where a bystander cannot be de-identified are discarded rather than delivered with a quality caveat. Site access and permission are confirmed for every location a contributor proposes.

The practical consequence is that legible screen and document content is not in the delivered pixels. If your pipeline needs the content of a document, raise it before the spec is signed; whether any engagement can keep specific content is not stated on this site.

Only partly, and the boundary is worth stating plainly. What is delivered is first-person video of a person at a desk, with hand pose, contact events and action segments for what the hands do: reaching for a phone, handling a keyboard, moving paper, picking up a pen. That is the physical, embodied side of desk work, and it is what a robot or a vision-language-action model needs.

The listed streams do not include a screen-capture stream, a cursor or click log, a keystroke log or an accessibility-tree dump, and legible screens are blurred at ingest. A model that must read an interface, or learn which click follows which state, is not served by this collection as described. We do not state any such stream, so nothing here should be read as one.

For a team that wants both, the sensible split is to treat office video as the embodied half of a pipeline and source interface traces separately. If you need synchronized screen content under a specific arrangement, say so in the request so it can be scoped rather than assumed.

Action segments are frame-accurate: each carries a start and end time, a verb and a noun, in the order the spec defines. Hand pose is 3D with a per-joint visibility flag, and contact events mark when a hand meets or releases an item, which is useful for tasks defined by a grip or a release such as feeding a sheet into a scanner or seating a stapler magazine.

Every episode passes calibration and synchronization checks on every active stream, plus label and annotation review for hand pose and action segments where those are included, before it is delivered. Episodes that fail are rejected rather than shipped with a caveat, and the reject log is delivered with a Spec Pilot batch so you can see what was refused and why.

Variation between people is expected rather than treated as a defect. Two contributors squaring a stack will not make the same motion, and the spec’s accept criteria define what counts as a correct completion, not a single reference movement.

Several things a buyer might assume are not stated on this site, and this page does not state them either: screen-capture, cursor, click and keystroke streams; a per-environment list of which of the ten streams ship; readable screen or document content in delivered video; task-completion or paper-state labels such as whether a document is correctly filed; force or tactile data; counts of offices, contributors, hours or episodes for the office environments; and a delivery timeline specific to offices.

The only timeline stated is the general one: a median of 48 hours from signed spec to first validated batch where coverage is deep, and three to four weeks where it is thin. Whether office coverage is deep or thin for a given location and task list is settled when the spec is scoped. No company, office or brand is named as a client or a capture location on this page.

Finished data ships in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or as raw streams with metadata, whichever your pipeline already consumes. RLDS and LeRobot are built around episodic, step-based sequences, which maps directly onto an office task already decomposed into task-unit episodes with action-segment labels. Because that structure is set at the spec stage, the same capture can be re-exported into another format later.

The license is perpetual, non-exclusive and buyer-owned by default, with exclusivity available, so the data can go into a commercial training run without a separate rights negotiation. Provenance and data-processing documentation accompany delivery, and the DPA covers the data-processing terms for the engagement.

Most programs start with a Spec Pilot, a small, fast first batch captured against the spec and QA’d against the same accept criteria as a full order. The same written spec then scales into production volume in batches.

Describe the tasks and workstations, the objects and starting states, the sub-steps of each task and their order, the office types and rooms, and the variation or failure cases your pipeline is missing. That becomes a written skill spec, quoted for reach, timeline and price before any capture begins, and most programs run a Spec Pilot first.