Industries / Retail, hospitality and personal services / Retail and e-commerce / Grocery
Checkout
Checkout work covers how shoppers record and pay for what they take: the scanner and the till, the ways of paying that are particular to food shopping, and the bags the goods leave in.
5
Milestones achieved, from 1995 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1995
Shoppers scan while shopping
In October 1995 a Finast supermarket in Lakewood, Ohio, owned by the Dutch grocer Ahold, gave 100 volunteer shoppers handheld scanners to scan goods as they picked them and then pay at a special station. Supermarket News called it the first such system live in a North American store and reported the idea already running in 26 Safeway stores in England and three Albert Heijn stores in the Netherlands. Volunteers agreed to random audits, in which staff check an order item by item. By May 2009 Stop & Shop and Giant ran a version, branded Scan It, in 250 stores, with nearly one million shopping trips a month.
LessonShopper scanning works by trusting most shoppers and checking a random few.
Credit
Finast and Ahold; Safeway (UK); Albert Heijn; Modiv Media and Motorola, for Scan It
2015
England charges for bags
Supermarkets had long handed out plastic carrier bags free at the till. From 5 October 2015 England required large shops to charge 5 pence for each single-use bag, raised to 10 pence and extended to all shops in May 2021. By the figures of Defra, the UK environment department, the main seven retailers sold 126 million single-use bags in the year to April 2026, against the 7.6 billion that the charity WRAP counted for 2014, a fall of 98%.
Credit
UK government (Defra)
2017
Cameras check self-checkouts
In 2017 Walmart began using software from Everseen, of Cork, Ireland, that watches ceiling-camera video at registers and self-checkouts and alerts staff when an item seems to go into a bag without being scanned. WIRED reported in May 2020 that it ran in thousands of Walmart stores. A group of Walmart corporate employees told WIRED that it often flagged honest shoppers and missed real theft. Walmart said false alerts were not a widespread issue and that changes had cut the number of alerts.
LessonNeither side published an error rate, so the dispute could not be settled from outside.
Credit
Everseen; Walmart
2019
Stores must take cash
Philadelphia passed a law in February 2019, in force from 1 July 2019, that requires almost all stores to accept cash, with fines of up to $2,000; New Jersey enacted a statewide rule in March. Lawmakers said cashless stores shut out people without bank accounts; the FDIC counted nearly 6% of Philadelphia-region residents as unbanked in 2017. In April 2019 Amazon said its Amazon Go stores, built to have no checkout at all, would start accepting cash. Massachusetts has long required shops to take cash.
LessonA store with no till still has to serve shoppers who pay in notes and coins.
Credit
Philadelphia City Council, with Councilman Bill Greenlee as sponsor; the New Jersey legislature, with Assemblyman Paul Moriarty as sponsor
2025
Chips on food-aid cards
US food aid (SNAP, called CalFresh in California) is spent at the grocery till with an Electronic Benefit Transfer (EBT) card, and California's cards carried only a magnetic stripe. Thieves fitted hidden skimmers to card readers, copied cards and drained them; in January 2024 California reimbursed about $20.9 million of stolen benefits. In February 2025 the state began issuing EBT cards with a chip that can also pay by tap, the first state to do so by its own account, and about 4 million were out by the end of April 2025. Shops' card terminals need a software upgrade to read them. The state says reported thefts fell 83% by November 2025; CalMatters reported that county fraud investigators think the state's count is too low.
LessonCalMatters reported in 2023 that the state had ignored warnings and delayed a proposal for chip cards.
Credit
California Department of Social Services; Office of Technology and Solutions Integration; FIS, the card processor
Watch for
Smart carts spread
Instacart's Caper Carts have basket-facing cameras, a scale certified by weights-and-measures authorities, location tracking and a touchscreen, so the cart keeps a running total as shoppers fill it. Instacart says the carts now run in more than 100 cities in 15 US states with more than a dozen retail banners, among them Kroger, Schnucks, Wakefern's ShopRite and Coles in Australia, that it tripled the number deployed in a year, and that an on-cart "Got everything you need?" prompt raised basket size by nearly one percentage point on average.
Achieved whenan outside party counts the carts in use and publishes how often each cart's item list matches what is in the basket.
Claimed by
Instacart, June 2026
Open
01
Name the loose produce
A public test of produce recognition at the scale: about 3,000 photos of loose fruit and vegetables taken from above a checkout scale, covering the 150 most-sold price look-up (PLU) codes, the four- or five-digit numbers on produce stickers. It should include look-alike varieties, such as Gala and Fuji apples, and produce in clear bags, and it should score several open models on the set.
The briefCredit · unclaimed02
A public missed-scan test
A labeled video set from a mock self-checkout, where volunteers ring up baskets of real products. Most transactions are honest; in some, items pass unscanned in common ways (two items stacked, a hand over the barcode, two items in one hand). Camera models are then scored on it.
The briefCredit · unclaimed
Open 01 · the brief
Name the loose produce
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of produce recognition at the scale: about 3,000 photos of loose fruit and vegetables taken from above a checkout scale, covering the 150 most-sold price look-up (PLU) codes, the four- or five-digit numbers on produce stickers. It should include look-alike varieties, such as Gala and Fuji apples, and produce in clear bags, and it should score several open models on the set.
Why it is still open
Checkout makers sell camera recognition for loose produce but publish no accuracy. Diebold Nixdorf says its own survey found shoppers take 10 to 15 seconds on average to pick one loose item from the on-screen list. Public datasets stop short: Klasson and colleagues (KTH, 2019) photographed produce on store shelves to help blind shoppers, and Teo and Chaw (Universiti Kebangsaan Malaysia, 2026) tested 26 classes of Malaysian produce. Searches of GitHub and Hugging Face found no set labeled by PLU code.
Built with what exists
A kitchen scale, a USB camera on an arm and LED lights; the PLU code list kept by the International Federation for Produce Standards (IFPS); OpenCLIP and an open vision-language model such as Qwen2.5-VL; Klasson's Grocery Store Dataset as a second test.
First step
Buy 20 kinds of loose produce, including three apple varieties, and photograph each kind 20 times on the scale under two lights.
Done looks like
3,000 labeled photos, 20 per PLU code, released under an open license. Published top-1 and top-3 accuracy for at least three models, overall and on the look-alike pairs. A second labeler agrees on at least 98% of a 300-photo sample.
Open 02 · the brief
A public missed-scan test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A labeled video set from a mock self-checkout, where volunteers ring up baskets of real products. Most transactions are honest; in some, items pass unscanned in common ways (two items stacked, a hand over the barcode, two items in one hand). Camera models are then scored on it.
Why it is still open
Walmart has used Everseen's cameras since 2017 to flag unscanned items. Walmart employees told WIRED in 2020 that false alerts were frequent, Walmart said they were not, and no error rate has been published. Searches of arXiv, GitHub and Hugging Face found no public video set of missed scans; the public checkout image sets label products, not scans.
Built with what exists
A mock checkout counter with a USB barcode scanner logging to a laptop; two overhead webcams; about 100 everyday products; CVAT for labeling video; video models such as Qwen2.5-VL or action-recognition models from PyTorchVideo.
First step
Build the mock counter, film 30 honest transactions, and line up each scanner beep with the video time stamps.
Done looks like
500 filmed transactions, about 10,000 item passes of which 5% are deliberate misses, each labeled by two people. For each model, the published share of misses caught and false alerts per 100 honest transactions, with the videos released.