Egocentric Data Collection by Industry

This site currently covers six industry settings for custom egocentric data collection: warehouse and logistics, retail and shelf operations, automotive service, tool use and assembly, kitchen and food preparation, and household chores. Each has its own dedicated custom-collection page describing the tasks it covers, and its own catalog dataset page showing example footage and dataset detail already captured in that environment.

This page is an index, not a replacement for those pages. If you already know your industry, the fastest path is its dedicated page; use this page to compare settings, see which page fits a task that spans more than one industry, or decide where to start if your use case does not map neatly onto a single setting.

This differs from the /solutions pages, which organize by data modality (audio, image, video, text, multimodal) rather than by industry setting, and from the general egocentric-video-data-collection and enterprise pages, which describe the capture process and program structure rather than enumerate specific industries.

Warehouse and logistics collection covers tasks inside working fulfilment space — pick aisles, pack benches, pallet building, and returns desks — rather than a mock aisle built for the camera. Typical tasks include single-item picking from bin to tote, barcode scanning, carton break-down, pallet stacking and shrink-wrap, label application, and two-person long-item carries.

The setting runs multi-person activity more than most other industries, since pickers, packers, and equipment share the same aisles, and sight-lines are often longer than other settings, which affects how depth is captured and flagged. The dedicated page covers the setting in full and links to the catalog dataset for example footage. For buyers who need the full receiving-to-loading flow rather than pick-and-pack alone, the logistics and fulfillment page covers receiving, putaway, picking, packing, sortation, and loading as one spec.

Retail and shelf-operations collection covers facing and front-stocking to a planogram, case-cutting and shelf-loading, price and label swaps, chilled-case rotation, and fixture resets on live retail floors — not a staged aisle. Customers are present throughout a live retail floor, which makes this one of the settings where bystander de-identification matters most: every bystander is blurred at ingest, and a clip is discarded rather than shipped if a bystander cannot be de-identified.

Because de-identification discards rather than downgrades, episode yield in retail settings tends to run lower than in a private setting like a kitchen — a tradeoff this site states plainly rather than smoothing over. The dedicated page and catalog dataset page cover the setting and example footage in detail.

Automotive service collection covers diagnostic, wheel, fluid, and trim work in working service bays — torque-wrench work to spec, fluid drain and fill, trim and clip removal, diagnostic connector seating, and filter swaps in constrained cavities. The setting is heavy on tool use, constrained reach, and frequent occlusion by the vehicle itself, and license plates and VIN plates are blurred at ingest without exception.

This is a setting with little public reference footage to compare against, so a written spec matters more here than almost anywhere else: the tasks, tool list, and bay layout a model needs to learn have to be described rather than assumed from an existing corpus. The dedicated page covers the setting and process in detail.

Tool use and assembly collection covers bench assembly, fastening, measuring, and fitting — driver and bit selection, clamp-align-fasten sequences, measuring and marking to tolerance, press-fit assembly, wire stripping and crimping, and tool hand-offs between hands. Because the operator controls the whole bench, this setting tends to give the cleanest capture for bimanual precision of any industry on this site.

Small fasteners and fine parts push below the practical resolution of some depth sensors, so hand pose carries more of the signal than metric depth on the smallest parts — a limitation this site states rather than papering over with uncalibrated depth. The dedicated assembly task data page covers a single named skill in detail; for a full multi-step production procedure captured as one sequenced workflow, with task decomposition and failure/recovery coverage across the whole run, see the manufacturing workflow data page.

Kitchen and food-prep collection covers two-handed food preparation in real domestic kitchens — knife work, pouring and decanting, opening jars and sealed packaging, loading and unloading a dishwasher, stovetop transfers, and bimanual dough handling. This is among the densest manipulation environments captured, with more contact events per minute than most other settings, which is why it is a common starting point for teams building manipulation-heavy pretraining data.

Wet and reflective surfaces degrade some depth-sensing approaches, so per-frame confidence information travels with the data rather than leaving a buyer to guess which frames are reliable. The dedicated kitchen page and the catalog dataset page cover the setting and example footage. For a commercial kitchen or restaurant operation specifically — prep lines, cooking stations, plating, and dish pits running at service pace — the dedicated food service page covers that narrower, faster-paced setting.

Household collection covers laundry, tidying, cleaning, and surface reset across whole homes — folding and sorting laundry, loading and unloading a washer, wiping and resetting surfaces, vacuuming with obstacle negotiation, bed-making, and bin changes. Episodes here run longer-horizon and lower-precision than a bench task, with more whole-home navigation between rooms than any other setting on this site.

Deformable objects — laundry, bedding, bags — dominate this setting, so instance masks are shipped per frame rather than propagated from a sparser set of keyframes, which is part of why annotation cost per hour runs higher here than in most other settings. The dedicated household page covers the setting and example tasks.

Construction and skilled trades is a thinner-coverage setting on this site than warehouse, retail, or kitchen capture, since site access and operator vetting take longer to arrange. The dedicated construction and trades page covers what a written spec for this setting names and how the capture, labeling, and privacy process applies to jobsite and trades work.

Because this is a thinner environment, first-batch timelines here typically run three to four weeks rather than the faster turnaround available in deeper-coverage settings — a timeline this site states up front rather than quoting the same number across every industry.

Clinical and care is one of the thinner-coverage settings on this site, and capture here is deliberately narrow: training environments with simulated patients only, never real patient footage. Within that boundary, the dedicated healthcare and hospital page covers non-clinical hospital-adjacent tasks — equipment handling, supply and cart handling, room turnover, and simulated patient-adjacent logistics — captured, labeled, and licensed the same way as any other custom collection.

Because this is a thin environment, first-batch timelines here typically run three to four weeks rather than the faster turnaround available in deeper-coverage settings, the same pattern as construction, agriculture, and hospitality.

Hospitality is one of the site's thinner-coverage settings, and it splits into two different kinds of capture: guest-facing, front-of-house work such as reception and concierge interactions, and back-of-house work such as cleaning and room-turnover tasks, which falls under the site's broader "cleaning & housekeeping" activity category. Front-of-house capture is bystander-dense, so yield after de-identification runs lower there than in less populated settings.

Because hospitality is a thin environment, first-batch timelines here typically run three to four weeks rather than the faster turnaround available in deeper-coverage settings, the same pattern as construction, clinical, and agriculture.

The stream set available — head/wrist RGB, depth, IMU, hand pose, and action segments — is the same menu across every industry; what differs is which streams a given task actually needs and how reliably each stream performs in that setting. A bench task benefits most from hand pose and depth at close range; a warehouse task benefits more from action segments and wider-range depth; a household task leans on navigation-friendly IMU and long-horizon action segmentation.

Privacy handling also differs by exposure, not by policy: every setting uses the same blur-at-ingest and discard-rather-than-downgrade approach to bystanders, but a retail floor or warehouse aisle has far more bystander exposure than a private kitchen or home workshop, so yield and the proportion of discarded clips vary by setting rather than by any difference in the privacy process itself.

Start with a written skill spec naming your industry setting — one or several — along with the tasks, scenes, objects, streams, and labels your training pipeline needs. That is the same intake format used for any custom collection on this site, whether the setting is a warehouse aisle or a kitchen counter.

From there, a Spec Pilot — a small, fast first batch captured against that spec — proves coverage, stream quality, and reject rate before you commit to full volume. Buyers that need more than one industry in a program, such as warehouse picking alongside bench assembly, write both settings into the same spec rather than running two separate engagements.