Cycle Counting and Inventory Check Data

Task anatomy

A cycle count checks a subset of locations against the inventory system on a rolling schedule, in place of a single annual stocktake. The worker, or a scanning device, goes to a location, identifies it, identifies what is stored there, counts it, compares the result with the system quantity, and either confirms or flags a difference. Flagged differences trigger a recount, a search for misplaced stock and an adjustment, often with a reason code.

The task combines perception with reconciliation. Perception covers reading a location label, recognizing the item or its case label, and counting units that may be stacked, partly hidden or in mixed pallets. Reconciliation covers comparing the count with the record and deciding what to do about a difference. A model that counts well but cannot reconcile has solved only half of the task.

Failure modes include unreadable or missing location labels, similar-looking items, partial cases, items stored in the wrong place, stock hidden behind other pallets, mixed lots and a system quantity that is itself out of date because of an unposted movement. This page belongs to the warehouse cluster of the humanoid data guide.

Why location structure matters

A location is a defined address in a layout: aisle, bay, level, position. How locations are assigned and numbered follows from storage assignment and layout decisions, which warehouse-order-picking research reviews alongside routing, batching and zoning (de Koster et al., 2007). For a counting dataset this means the same image can mean different things depending on the site's numbering scheme and pick-face type. Ask for the location scheme alongside the images; see rack and shelf location images and pick face types.

Which data helps which part

Retailers and third-party logistics operators hold count history, and some also hold images taken during counts or audits. Scanning programs may have captured images at scale for other purposes, such as shelf compliance. Check how they were taken before treating them as counting data.

What the data can support

• Counting and recognition. Detection of stacked cases, shelf facings and pallet layers, with verified counts as labels.
• Location reading. Rack label recognition and mapping to a layout.
• Exception reasoning. Language and rules linking variance patterns to likely causes, such as misplaced stock or unposted moves.
• Evaluation. Comparing a system's count and variance flag against audited counts on held-out locations and sites.

What this data cannot support

• Robot actions. Images and records contain no joint states, motion commands or contact data. A humanoid moving to a rack, opening a tote or lifting a case needs robot-specific observations and actions.
• Hidden stock. An image shows what is visible. Counts of stock hidden behind other loads rely on the human's knowledge, which the data may not record.
• Count ground truth by default. A count record is the worker's claim. Audited recounts are stronger evidence; ask which are available.
• Client-confidential product data. Item numbers, quantities and customer names belong to the client and cannot be passed on without permission.
• Viewpoint transfer. Images taken at human eye height or from a scanning device differ from a robot's sensor positions. Camera pose and calibration matter, and archives often lack them.

Illustrative record: image, count and reconciliation

# Illustrative only - synthetic values, not an available dataset
count_id: cc_000812
site: dc_044
location: {code: "A-14-03-2", scheme_version: 3, pick_face: pallet_slot}
images:
  - {id: img_5521, mount: handheld, height_m: 1.5, lighting: warehouse_led, camera_pose: unknown}
item: {item_ref: item_hash_77a1, case_label_visible: true}
counts:
  system_qty: 48
  counted_qty: 45
  counter: operator_A12
  recount: {done: true, qty: 45, by: operator_B03}
variance: {units: -3, reason_code: misplaced_stock, resolved: found_in_B-02-01}
labels: {stack_layers: 3, partial_case: true, occlusion: moderate}

Fields like camera pose, scheme version and recount status decide whether a record can be used for a robot's perception.

Sourcing and permissions

The route is existing assets: ask a third-party logistics operator, distributor or retailer for count histories and any images attached to them. If images are absent or unsuitable, a purpose-made capture is a separate decision; see existing footage versus commissioned recording. Related references are barcode scanning workflows and warehouse layout and facility data.

Inventory belongs to clients in a 3PL setting, and a retailer's counts may reveal stock levels that are commercially sensitive. Ask whether clients allowed use, whether images show people or customer branding, and what the contracts say about reuse; see client-owned inventory rights. Redaction does not on its own settle permission. This is not legal advice.

Quality and acceptance checks

Buyers set the thresholds. Checks worth writing into an acceptance test include the following.

• Count reliability. Take a sample of locations and have them recounted independently; compare with the supplier's counts and document the mismatch rate.
• Image sufficiency. For each sampled record, check that a person could reach the recorded count from the images alone. If not, the image cannot serve as the label's evidence.
• Variance mix. Check that variances of different sizes and reasons are represented, not only zero-variance counts.
• Lighting and height spread. Confirm images cover the upper rack levels and low light, where recognition fails first.
• Held-out design. Split by site and by period, since a location scheme and packaging change over time.
• Documentation. Ask for a datasheet that records how images and counts were collected. Datasheets for Datasets is a useful template.

Questions to ask suppliers

1. How many locations, sites and rack types are covered, and what is the location numbering scheme?
2. Were images captured during counts, and with what device, height and lighting?
3. Is camera pose or calibration recorded for any images?
4. Are counted quantities first counts, recounts or audited counts, and are they marked?
5. What share of counts show variances, and are reason codes standardized?
6. Which item and client fields are confidential, and how are they pseudonymized?
7. Do images contain people, screens or customer branding, and how are they handled?
8. Which clients permit use of their inventory records and images for model training?
9. Can count history be split by site and time so evaluation sets can be held out?

Discuss your requirements

If your project needs counting or inventory-check data, describe the site type, modalities and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match, and nothing is contracted until a supplier agrees. See also the catalog entry for images and inspection photos.

Explore sourcing and licensing options with SourceX

Sources

• Design and control of warehouse order picking: A literature review — Literature review covering layout design, storage assignment and routing, which shape how locations are structured and counted.
• Datasheets for Datasets — proposes datasheets documenting dataset composition and collection process.