Grocery Bagging and Store Order Assembly Demonstrations

The buyer's problem

Bagging looks trivial because people do it without thinking. It is a planning task disguised as a manipulation task. A cashier or order picker sorts items by weight and fragility, keeps cold goods together, puts eggs on top and cans at the bottom, opens a bag whose handles flop, and judges when one more item will tear it. The bag is a soft container whose shape changes with every placement.

Two store workflows matter: staff bagging at a checkout, and assembling a click-and-collect order from shelves into totes and bags. They share the packing logic and differ in the surroundings. The first is a fixed counter; the second is mobile, travels aisles and uses a pick list. This page covers both as demonstrations. The shelf side is described in retail shelf restocking.

Why checkout datasets do not solve it

Retail vision work is largely about recognizing what is on the counter. The RPC dataset, for example, contains single-product images captured under controlled conditions and multi-product checkout images, labeled at several levels across 200 product categories grouped into 17 meta-categories, with up to 83,000 images (Wei et al., 2019). It supports automatic checkout, which is an identification and counting problem.

That is object recognition, not manipulation training. Knowing that an item is a bag of flour does not tell a robot how to lift it without tearing the pack, how to orient it in a soft bag, or when it will crush the bread beneath it. For the general distinction, read object recognition data vs manipulation training data. Catalog data can help the perception layer; see product catalog data for robot perception.

Task anatomy

Bagging at a counter

1. Open or place the bag, check its type (paper, plastic, reusable).
2. Pick items from the belt or the scan area, reading weight and fragility cues.
3. Decide order: heavy and rigid first, fragile and crushable last, cold and frozen grouped, non-food apart from food.
4. Place items so the bag keeps its shape, and close or hand over.
5. Start a new bag when the first is full.

Order assembly in store

1. Read a pick list or app prompt and walk to a location.
2. Pick the correct variant, check expiry or size, and handle substitutions.
3. Place items in a tote by temperature zone and weight.
4. Stage the order for pickup; swap, tag or label it.

Failure modes worth capturing on purpose are a bag tearing, an item rolling out, eggs or glass broken, wrong variant selected, a frozen item thawing in the wrong tote, and a person asking the worker something mid-task.

What helps and what does not

Overhead point-of-sale cameras cover the counter and may show items, but not the bag's interior. A chest or head-mounted camera sees more of the packing action. Combine them, as set out in head-mounted camera capture specification.

Sourcing

Existing footage. Grocers have loss-prevention and checkout cameras. These capture customers, payment moments and staff, and were installed for security and shrink control, not AI training. Reusing them raises purpose and notice questions that the grocer, not the buyer, must answer; this is not legal advice. Expect a hard reuse restriction on anything that shows customers. Training videos for baggers are better behaved but are short and idealized.

Commissioned sessions. A store can run sessions in a closed lane or before opening, with staff volunteers and purchased or donated stock. You control product mix, order lists and views, and you can add fragile and heavy items on purpose. The mix of sealed packaged goods, produce, bottled drinks, bread, eggs and frozen items matters more than item count. This is where fragile item handling and deformable object handling become relevant.

If you need a mixed approach, describe the store type and packing tasks to SourceX and it will investigate whether US businesses hold suitable material or whether a commissioned session is worth scoping.

Illustrative package

# Illustrative only - synthetic values, not an available dataset
package: bagging_pilot
site: grocer_site_022, closed checkout lane, off-hours
orders: 60 synthetic baskets, 8 to 25 items each, composed from a fixed list of 120 SKUs
product_mix: [cans, boxed dry goods, produce, eggs, bottled drinks, bread, frozen, glass jars]
bags: [paper, plastic, reusable_cloth]; fill state logged
workers: 6 (bagger_E01..E06), written consent
views: head_mounted; counter_overhead; bag_interior_close_view
metadata_per_item: sku_id, weight_class, fragility_class (low/med/high), temperature_zone
labels: item_order, placement_zone, outcome (intact / crushed / torn / retried)
excluded: grip forces, bag tension, robot actions
rights: training and evaluation; no redistribution

Acceptance pointers

• Basket realism. Compare item mix against a real store, and check that heavy and fragile pairings occur often enough to show the packing rule.
• Label independence. Fragility classes should come from a written scheme, not from the demonstrator's impression.
• Outcome coverage. Torn bags and crushed items should be present. A dataset with no failures cannot show recovery.
• Customer-free frames. Verify no customer faces, loyalty cards or payment screens appear, and ask what masking removed.
• Held-out items. Reserve SKUs and a store layout not seen in training. See success and outcome labels.

What this data cannot support

• Robot packing behavior. Human demonstrations contain no robot actions, grasp commands or compliance control. A robot needs its own grasp and placement data on soft containers.
• Mass and force. Weight, grip pressure and the squeeze that crushes a loaf are invisible. Log weights; do not infer them.
• Soft-bag physics. Simulating a deformable bag with accuracy is difficult, so published-style validation on a real bag stays necessary.
• Customer interaction. A cashier asks about preferences, handles refusals and manages queues. Staged data omits most of this.
• Catalog changes. Packaging redesigns, seasonal stock and promotions shift appearance faster than a dataset updates.

Questions to ask suppliers

1. Were customers in frame at any point, and how was this handled?
2. What was the purpose of the original cameras, and does that purpose limit reuse?
3. Does the footage show inside the bag, or only the counter?
4. How many distinct SKUs, bag types and basket compositions appear?
5. Are weights, fragility and temperature zones logged, and by whom?
6. Are torn bags, crushed items and re-bagging kept?
7. For order assembly, how are pick lists captured and is customer data removed?
8. Do pickers walk full aisles on camera, or is only the staging area recorded?
9. What consent did staff give, and does it cover model training?
10. Can a store or product set be reserved for evaluation?

Discuss your requirements

Discuss your humanoid training-data requirements with SourceX by describing the store type, task, views and permitted uses. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match, and nothing is contracted until a supplier agrees. See also the workplace task demonstration hub, the humanoid training data guide, and the catalog pages for product catalogs and descriptions and video recordings.

Describe the workplace demonstration data your robotics project needs

Sources

• RPC: A Large-Scale Retail Product Checkout Dataset — checkout recognition data focused on identification across product categories.