Industries / Retail, hospitality and personal services / Retail and e-commerce / Grocery
Shelves and stock
Shelf and stock work keeps a grocery store's shelves full and its aisles safe: ordering for each store, moving goods from the back room to the shelf, finding gaps, and clearing spills.
5
Milestones achieved, from 1970s to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1970s
Handhelds place store orders
Telxon, of Akron, Ohio, began designing order-entry and inventory systems for grocery, drug and hardware retailers in the early 1970s, and demand grew after the grocery trade adopted the UPC barcode as its standard in 1973. Its portable terminals took order and stock entries in the store and sent them over telephone lines, which cut out retyping and the errors that came with it. By 1993, according to a company history, 24 of the top 25 US grocery chains used Telxon's portable computers, for jobs such as receiving deliveries, managing prices and computer-assisted ordering.
Credit
Telxon Corporation
2004
Automated depots, empty shelves
In October 2004 J Sainsbury, the UK grocer, said the new logistics system from its four-year, £3 billion modernization program could not track stock properly, and it hired 3,000 extra shelf stackers to fill the gaps by hand. The Register reported write-offs of £140 million for IT and £120 million for equipment in its automated depots, plus £30 million of stock lost or damaged. Sainsbury's blamed Accenture, which ran its IT; Accenture said the depot automation was not part of its contract.
LessonNew depots and software did not keep shelves full while the stock records were wrong, and the gap was closed with people.
Credit
J Sainsbury plc; Accenture
2019
A robot hunts spills
In January 2019 Ahold Delhaize USA began putting Marty, a tall wheeled robot from Badger Technologies, into nearly 500 Giant, Martin's and Stop & Shop supermarkets to look for spills and fallen goods. It steers with lidar, a laser range sensor, and carries eight cameras. The Associated Press reported that the robot stops when it sees a possible spill, and people in a control center in the Philippines check the images before it calls staff over the loudspeaker. On the day reporters watched, it flagged a fallen price tag and a sprig of herbs.
LessonThe robot found hazards by camera, but a person still confirmed each alert.
Credit
Badger Technologies, a division of Jabil; Ahold Delhaize USA
2020
Panic buying empties shelves
When COVID-19 reached the US in March 2020, shoppers stockpiled food and household goods for about two months. NielsenIQ, a market research firm, found that out-of-stocks lingered in US supermarkets for the rest of the year. From May 2020 to February 2021 its 10 worst-hit categories lost $2.93 billion of sales because goods were not on the shelf, led by bath tissue at $836.56 million and paper towels at $689.29 million. By NielsenIQ's survey, 70% of shoppers buy another brand when an item is missing and 30% go to another store.
LessonShortages outlasted the buying spree by many months.
Credit
NielsenIQ, On Shelf Availability Barometer
2025
Cameras watch the shelves
Focal Systems makes small cameras that sit on the shelf edge, photograph the shelf opposite at set times and flag empty or low facings to staff. Walmart Canada put them in all its stores in 2022 after a 70-store pilot, with alerts fed into its existing inventory system. In 2025 Focal fitted more than 200,000 cameras in 498 Morrisons supermarkets in the UK in eight months, and there the cameras also reorder stock automatically. The trade magazine The Grocer gave the project its 2025 technology award; its write-up says availability rose by more than 2 percentage points across the 498 stores, a figure from the entry, not from an outside count.
Credit
Focal Systems; Walmart Canada; Morrisons
Watch for
Shelf maps from delivery shoppers
Instacart says it collects more than 10 million data points a day from about 600,000 of its shoppers, who scan shelves while picking customers' orders in stores across North America. It combines these with data from 1.6 billion orders and with images from its Caper smart carts to show grocers and brands, store by store, which products are out of stock. It is rolling the software out with "select grocers" first.
Achieved whenan outside party compares Instacart's out-of-stock calls with shelf checks made in the same stores at the same hours, and publishes how often they agree.
Claimed by
Instacart, September 2026
Open
01
Spill alerts tested on video
A public set of video from store-style ceiling cameras that shows staged spills (water, cooking oil, rice, a dropped jar of colored water) among many hours of ordinary aisle traffic, with every spill labeled, and a scored comparison of spill detectors on it: how many spills each finds, how fast, and how many false alerts it raises per camera-hour.
The briefCredit · unclaimed02
An open shelf camera
A low-cost camera, clipped to a shelf edge, that photographs the shelf opposite every 15 minutes and sends staff a list of empty facings, with open hardware and software, run for a month in one independent grocery store that agrees to the trial.
The briefCredit · unclaimed
Open 01 · the brief
Spill alerts tested on video
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of video from store-style ceiling cameras that shows staged spills (water, cooking oil, rice, a dropped jar of colored water) among many hours of ordinary aisle traffic, with every spill labeled, and a scored comparison of spill detectors on it: how many spills each finds, how fast, and how many false alerts it raises per camera-hour.
Why it is still open
Researchers at Clarkson University wrote in 2020 that they knew of "no publicly available dataset for supermarket hazards." They built their own from 1,180 real still images plus synthetic ones and scored image classification, not alerts over hours of video. In stores, Marty's spill calls were checked by people in the Philippines before staff were called. Searches of arXiv, GitHub and Hugging Face found spill datasets only for oil at sea.
Built with what exists
Two or three IP cameras or phones on tall tripods; a mock aisle of shelving, or a store aisle after hours with the manager's consent; Ultralytics YOLO segmentation models; an open vision-language model such as Qwen2.5-VL as a second check on each alert; the EdgeLite code the Clarkson team published on GitHub; for footage from a moving camera, Ring's Rover Cam, a camera on wheels that ships in the US from 16 December 2026 (see the Note "A camera on patrol").
First step
Film 20 staged water and rice spills on a mock aisle floor under two light levels, and label when each spill starts.
Done looks like
200 staged spills of five kinds and 100 camera-hours of ordinary traffic, labeled by two people. For each detector, the published share of spills found within 60 seconds and the false alerts per camera-hour.
Open 02 · the brief
An open shelf camera
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A low-cost camera, clipped to a shelf edge, that photographs the shelf opposite every 15 minutes and sends staff a list of empty facings, with open hardware and software, run for a month in one independent grocery store that agrees to the trial.
Why it is still open
Shelf-edge cameras run in Morrisons' 498 stores and in all Walmart Canada stores, but they are Focal Systems' closed product, and their results come from the retailers, such as Morrisons' availability gain stated in an awards entry. Research models and open-source projects on GitHub find empty shelf space in still photos, for example OOS-DSD (2025) and YOLO-based detectors, but none found reports a trial in a working store or a cost per meter of shelf.
Built with what exists
Raspberry Pi Zero 2 W boards with Camera Module 3, or ESP32-S3 camera boards; battery packs; an open detector such as Ultralytics YOLO trained on public empty-shelf images; a web page or chat message that lists gaps for staff; faces blurred before any image is stored.
First step
Mount one camera opposite a pantry or shop shelf, photograph it every 15 minutes for a week, and mark the gaps by hand.
Done looks like
20 cameras covering 40 meters of shelving for four weeks. Each alert is checked against a staff walk at the same hour. Published: the share of real gaps flagged, the share of alerts that were real, minutes from gap to alert, and hardware cost per meter of shelf.