Industries / Retail, hospitality and personal services / Retail and e-commerce
Shelves and stock
Stock work keeps the right goods on store shelves: counting stock, reordering, finding gaps, refilling shelves and tracking losses from theft, damage and errors.
7
Milestones achieved, from 1954 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1954
Factories copy the supermarket shelf
Toyota introduced a "supermarket method" in 1954. As in a supermarket, a worker takes only what is needed, and only what was taken is refilled. It grew into the kanban card system and Just-in-Time production. Toyota's own history says the idea came from a 1954 US trade article about a Lockheed aircraft plant, at a time when "supermarkets were not common in Japan."
LessonThe supermarket's way of refilling only what was taken was clear enough for factories to copy.
Credit
Toyota Motor Co., which credits Taiichi Ohno with the basic framework of Just-in-Time
2002
Empty shelves measured worldwide
A worldwide study of fast-selling consumer goods by Gruen, Corsten and Bharadwaj found that 8.3% of items were out of stock on average, about one in every 13 items a shopper wants. About three-quarters of the causes were inside the store: 47% came from store ordering and forecasting, and 25% were goods in the building but not on the shelf. The authors noted that the figure had not changed since a 1996 study.
LessonShelf availability is mainly a store-process problem.
Credit
Thomas Gruen, Daniel Corsten and their co-author Bharadwaj
2005
RFID trial cuts empty shelves
Walmart's top suppliers put RFID tags, chips that a reader detects by radio without needing to see them, on cases and pallets. A 29-week University of Arkansas study, commissioned by Walmart, compared 12 stores receiving tagged cases with 12 stores without them and found 16% fewer out-of-stocks on tagged items, which the researchers revised to about 30% in 2006. The tags told staff when cases in the back room should go to the shelf, so tagged items were refilled three times as often.
LessonThe gain came from a changed routine, an automatic pick list for the back room, not from the tag alone.
Credit
Bill Hardgrave and Matt Waller, University of Arkansas RFID Research Center; Walmart
2008
Most stock records are wrong
Nicole DeHoratius and Ananth Raman examined nearly 370,000 stock records from 37 stores of one retailer and found 65% did not match what was really there. Automatic reordering trusts those records. Goods the system wrongly believes are on hand, called phantom inventory, do not get reordered.
Credit
Nicole DeHoratius (University of Chicago) and Ananth Raman (Harvard Business School), in Management Science
2020
Walmart drops shelf robots
In January 2020 Walmart planned to put Bossa Nova's shelf-scanning robots in 1,000 of its 4,700 US stores. By November it had ended the contract, with robots in about 500 stores. The robots could flag gaps and wrong price labels, but "these robots cannot restock items," and workers already walking the aisles to pick online orders could update stock as they went.
LessonFinding a gap is only half the job, and workers already on the floor can sometimes do it more simply.
Credit
Walmart; Bossa Nova Robotics
2022
RFID tags spread beyond clothing
After tagging clothing item by item, Walmart required suppliers to put RFID tags on home goods, sporting goods, electronics and toys by September 2022, and on more than a dozen further categories by February 2024. Walmart says the tags improved its stock accuracy but has published no figure.
Credit
Walmart
2023
A shrink statistic withdrawn
Shrink is the gap between the stock a retailer's records say it has and what a physical count finds. The NRF put average US shrink at 1.6% of sales in fiscal year 2022, about $112.1 billion. An April 2023 NRF report said "nearly half" of 2021 shrink came from organized retail crime. On 1 December 2023 the NRF removed that claim after Retail Dive found it rested on the NRF's own survey from 2016.
LessonShrink can be measured in total, but its causes are hard to pin down, and a weakly sourced split spread widely before anyone checked it.
Credit
NRF; Retail Dive
Watch for
Shelf-scanning robots at scale
Simbe Robotics says it has more than 3,000 Tally shelf-scanning robots under contract with over 75 retail chains in nearly a dozen countries, and its website claims "60% fewer out-of-stocks" and "90% fewer pricing errors."
Achieved whenan independent study compares shelf availability and price accuracy in stores with and without the robots.
Claimed by
Simbe Robotics, September 2026
Drink restocking with less help
Telexistence's robots refill convenience-store drink coolers from the chilled back room, with remote operators taking over for error cases such as fallen bottles. In June 2025 Telexistence partnered with Physical Intelligence to automate more of the task with AI models, and in September 2025 it began pilots at some 7-Eleven stores in Tokyo.
Achieved whenan outside party reports robots refilling coolers in routine use, and how many tasks still need a remote operator.
Claimed by
Telexistence, June and September 2025
Humanoids restock a Danish supermarket
Hive Robots began training Unitree humanoid robots in June 2026 at a Føtex supermarket outside Copenhagen, owned by Salling Group, to collect products and stock shelves after closing. "Human operators still guide the robots through each step," and the aim is "autonomous shelf restocking by early 2027."
Achieved whenthe robots restock unattended in routine use, and an outside party reports how much they stock and how often people step in.
Claimed by
Hive Robots and Salling Group, October 2026
Open
01
Test the app's in-stock claims
An independent audit that compares what a retailer's app or website says is "in stock" at a named store with what is on that store's shelves at the same hour, published chain by chain.
The briefCredit · unclaimed02
Robot fronts a working aisle
A robot that "fronts" shelves, pulling products forward and lining them up so the shelf looks full, in a working shop after closing time, with no one steering it.
The briefCredit · unclaimed
Open 01 · the brief
Test the app's in-stock claims
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent audit that compares what a retailer's app or website says is "in stock" at a named store with what is on that store's shelves at the same hour, published chain by chain.
Why it is still open
The best-known figures are polls. In 2023 Retail Insight, which sells inventory software, polled over 1,000 UK shoppers, and 32% said items had been "listed as in-stock online" but were not there when they arrived. That measures what shoppers remember, not what the shelves held. A 2008 study of stock records found 65% of them wrong at one retailer. A search found no published audit of app claims against shelves.
Built with what exists
each retailer's public store-availability pages; Playwright to save what the app says, with a time stamp, just before the shelf check; a phone camera for the shelf; Python.
First step
Pick 50 everyday items and check them against the app in two stores of one chain.
Done looks like
2,000 item checks in 40 stores of four chains, each looked up within 30 minutes of the shelf check, with a second person rechecking 10% of shelves. The share of "in stock" claims with no item on the shelf is published for each chain.
Open 02 · the brief
Robot fronts a working aisle
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A robot that "fronts" shelves, pulling products forward and lining them up so the shelf looks full, in a working shop after closing time, with no one steering it.
Why it is still open
A 2025 paper by Davide Peron and colleagues (University of Padova, KTH and Animum) reports a Hello Robot Stretch 3 stocking and fronting in "a mock supermarket environment set in our laboratory," with over 98% success across more than 700 stocking events. The authors write that adoption in retail "remains limited." The humanoids in the Føtex trial (see Watch for) still need operators to guide each step.
Built with what exists
a Hello Robot Stretch 3, the commercial robot the paper used, with ROS 2; or an SO-101 arm on a rolling cart with LeRobot and a robot model trained from a few demonstrations, such as NVIDIA GR00T 2 (see the Note "Five demonstrations per task"); webcams.
First step
Agree with a local independent shop to film a worker fronting one aisle, and record 50 demonstrations on a copy of one shelf.
Done looks like
20 filmed overnight runs on a 10-meter aisle section of a working shop, with a published tally of facings fronted correctly, items dropped or knocked over, and human interventions, set beside a worker's time for the same section.