Egocentric Video Data for Logistics and Fulfillment

Logistics and fulfillment covers the end-to-end flow inside a distribution center or 3PL facility: dock-to-stock receiving, putaway to a storage location, single-item and case picking, pack-out, conveyor and tilt-tray sortation, and dock loading. A model that only sees one station in isolation has no exposure to how work actually moves between them.

Public corpora that touch logistics are largely incidental footage of one station rather than a spec-driven capture of the full flow, and they carry no hardware-synchronized depth or manipulation-grade hand pose in the first place. A custom collection written against the full flow fills that gap directly.

Training station by station without a shared spec also means each station can end up with a different camera placement, a different label set, or a different accept bar, which makes it harder to combine batches into one consistent training run later. Writing receiving, putaway, picking, packing, sortation, and loading into a single spec up front keeps the data comparable across stations from day one, even if a buyer only captures a subset of them first.

A fulfillment spec typically names tasks station by station: receiving (unload, dock-to-stock scan, damage check), putaway (location scan, shelf or bin placement), picking (single-item pick from bin to tote, case pick, barcode scan), packing (carton break-down, pack-out, label application), sortation (conveyor feed, tilt-tray or diverter handling, chute sorting), and loading (pallet build, shrink-wrap, trailer loading sequence).

A buyer does not need every station covered in one order — a spec can name just picking and packing, or the full receiving-to-loading flow, depending on what the training pipeline needs next. Coverage is scoped to the written spec, not assumed.

A head-mounted camera keeps the task motion in frame from the viewpoint a logistics policy will actually see from, following a worker through dock, aisle, pack bench, sortation line, and loading dock across a shift. Depth, hand pose, and action segments are aligned to video on one hardware clock.

Per-session recalibration accounts for cabling routed under hi-vis PPE, which shifts wrist extrinsics slightly between sessions. Long sight-lines down a pick aisle or across a sortation line push some pixels beyond reliable depth range; those are flagged rather than interpolated, so a buyer knows exactly where the depth stream stops being trustworthy.

Because a worker moves between stations during a shift, action segments also mark the transition points — leaving a pick aisle, arriving at a pack bench, feeding a sortation line — so a buyer can split a single continuous recording into per-station episodes during training without re-reviewing raw footage.

Delivered episodes carry 3D hand pose through reach, grasp, and place; instance segmentation and action segments marking task boundaries; and grasp and contact events. Depth ships synchronized to video, with distance-invalid pixels beyond reliable range flagged rather than filled in.

Every batch is measured against the accept criteria written into the spec before anything ships, billed per validated hour, with a reject log delivered alongside the data. Rejected items are recaptured at no cost to the buyer.

Each episode links to a signed contributor release through a documented consent chain. Faces, license plates on yard vehicles, and screens are detected and irreversibly blurred at ingest, and a reviewer confirms the result before anything ships; the anonymization record is delivered with the data.

A facility running at multi-person density, with pickers, packers, and equipment sharing the same aisles and dock space, means more bystanders pass through frame than in a lower-traffic setting. The same blur-at-ingest policy applies regardless of traffic; a clip that cannot be de-identified is discarded rather than shipped with a caveat.

Every custom collection ships under a perpetual, non-exclusive, buyer-owned commercial license by default; exclusivity is available on request so the data is never resold to another buyer. The consent chain and anonymization record travel with the delivery.

This is the same licensing and documentation structure used across every custom collection on this site — nothing fulfillment-specific changes about ownership or rights.

Logistics and fulfillment data ships in RLDS, LeRobot, and WebDataset, plus HDF5, zarr, and Rerun .rrd for inspection and debugging. Converter source is readable, so a buyer can retarget the schema to a different loader if needed.

These are the same delivery formats available across every custom collection on this site, chosen to drop into common imitation-learning and robotics training pipelines without a bespoke conversion step.

Describe the stations and tasks your training pipeline needs — picking only, packing and sortation, or the full receiving-to-loading flow. That becomes a written spec covering tasks, environments, capture tier, annotation depth, and accept criteria, quoted for reach, timeline, and price before any capture begins.

A Spec Pilot — a small, fast first batch run against that spec — proves coverage and reject rate on your facility type before you commit to full volume. On-demand top-ups run against the same spec and accept criteria afterward, so new data stays consistent with what you already hold.