Industries / Transport and vehicles / Warehousing and distribution
Counting stock
Counting stock is how a warehouse knows what it holds and where: the computer record for every location, and the counts that check that record against the shelf.
7
Milestones achieved, from 2003 to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2003
Tag mandate, retreat, return
In June 2003 Walmart told its top 100 suppliers to put radio-frequency identification (RFID) tags, chips that a reader detects by radio without a line of sight, on pallets and cases, starting in early 2004, and widened the rule to more suppliers later. Suppliers carried the cost and some complained that Walmart was not using the data, and Walmart backed away. It came back at item level, with RFID on apparel in its stores from 2020 and on home goods and other departments from September 2022, by when tag prices had fallen by up to 80%, to about 3 to 5 cents (McKinsey figures, as reported by RFID Journal).
LessonA Gartner analyst said the first mandate failed because there were no standards and tags were costly; it held once tags were cheap and the data served more than one job.
Credit
Walmart; its suppliers, including Procter & Gamble
2008
Most records are wrong
Nicole DeHoratius and Ananth Raman checked nearly 370,000 inventory records from 37 stores of one retailer and found 65% were inaccurate: the system's count did not match what was on hand. Regular audits reduced the errors, and complex stores and the way goods were distributed increased them.
LessonPlanning software that assumes the record is right plans from a wrong count.
Credit
DeHoratius (University of Chicago) and Raman (Harvard Business School), in Management Science
2018
Tags expose shipping errors
A 12-month pilot followed RFID-tagged items from eight brands through their distribution centers to five retailers. For 69% of inbound orders, what the brand said it shipped did not match what the retailer believed it received. The lab cites earlier studies putting store stock records at about 63% accurate without RFID and about 95% with it.
Credit
Auburn University RFID Lab and GS1 US (Project Zipper)
2020
Walmart drops shelf robots
Walmart ended its contract with Bossa Nova Robotics, whose camera robots had scanned shelves for empty and misplaced items in about 500 stores. The Wall Street Journal reported that Walmart found people could get similar results.
LessonA counting robot has to beat the simpler option of staff doing the check, not just do the check.
Credit
Walmart and Bossa Nova Robotics
2023
Drones count at IKEA
Ingka Group, the largest IKEA retailer, said 100 autonomous drones from Verity, a Swiss company, were checking stock in 16 locations in seven countries, flying outside opening hours. The first flew in IKEA Switzerland in 2021. By August 2024 IKEA reported more than 250 drones at 73 locations in nine countries.
Credit
Verity; Ingka Group and Inter IKEA as customers
2025
GXO adopts scanning robot
After a pilot, GXO said it would deploy Dexory's ground robot, which scans racks up to 13 meters high, at its Oosterhout, Netherlands, site and later in the United States and the rest of Europe. GXO said the system "eliminated manual stock checks and cycle counts" there; a cycle count checks a few locations each day instead of shutting the building for a yearly count. Dexory's own figure is 10,000 pallets scanned an hour.
Credit
Dexory (UK); GXO as customer
2026
Drones make counts monthly
ASICS America flies Corvus One drones seven times a night in the very narrow aisles of its Byhalia, Mississippi, distribution center, which holds about 3 million units. Two high-risk storage areas used to be counted twice a year and are now counted monthly. ASICS says it finds misplaced stock about 60% faster and has moved about 200 hours a month of counting work to fixing errors; on each pass the drone reads every product barcode and also counts the boxes, so the two totals can be checked against each other.
Credit
Corvus Robotics; ASICS America as customer
Watch for
Longer-range rack scanning robot
Dexory launched a scanning robot with a range of up to 60 feet, up from 40 feet, and a feature called Storage Health that flags damaged racking, broken pallets and unstable loads in its images. Dexory says its customers have cut audit time by 80% and raised throughput by 20%, but does not name them.
Achieved whenA named customer publishes those figures, or an independent check compares Storage Health's flags with a manual rack inspection.
Claimed by
Dexory, February 2026
Open
01
Camera counts against hand counts
A test racking bay of about 100 locations holding labeled cartons, whose true contents are fixed by two independent hand counts. A third person moves cases and changes quantities on purpose. Then a camera system photographs every location, and software counts the boxes and reads the labels. The camera can be a phone on a pole, a small indoor drone flown in the empty bay, or a ground robot that drives itself to saved points (see the Note "A camera on patrol"). Error rates are published by type: wrong item, wrong count and missed location.
The briefCredit · unclaimed02
Labels read where they hang
A public set of photos of pallet and case labels taken where they sit in racking: from the floor, at heights up to the top beam, at angles, in poor light, torn, wrinkled or behind shrink-wrap glare, each with its true item number, lot, quantity and case count. The project publishes how often a vision-language model reads each field correctly, and how many of its errors a second reviewer model flags for a person to check.
The briefCredit · unclaimed
Open 01 · the brief
Camera counts against hand counts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test racking bay of about 100 locations holding labeled cartons, whose true contents are fixed by two independent hand counts. A third person moves cases and changes quantities on purpose. Then a camera system photographs every location, and software counts the boxes and reads the labels. The camera can be a phone on a pole, a small indoor drone flown in the empty bay, or a ground robot that drives itself to saved points (see the Note "A camera on patrol"). Error rates are published by type: wrong item, wrong count and missed location.
Why it is still open
Published figures come from vendors and customers, and they report speed and labor saved, not error rates against a full hand count. Gather AI cites "900 pallets per hour with 3 drones" against customers' estimate of 60 bins an hour by hand, and ASICS reports faster recovery and fewer hours but no error rate. No public controlled test turned up.
Built with what exists
Ordinary steel shelving; printed Code 128 barcode labels; a phone camera; an open-source barcode reader (ZXing or ZBar); a vision-language model to count boxes (see the Note "Three working devices").
First step
Set up 20 locations, plant five errors, and run one round.
Done looks like
Ten rounds over 100 locations with planted errors, with the share of planted errors caught and the false alarms, each with a 95% confidence interval, beside the time taken for a timed hand count of the same bay.
Open 02 · the brief
Labels read where they hang
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of photos of pallet and case labels taken where they sit in racking: from the floor, at heights up to the top beam, at angles, in poor light, torn, wrinkled or behind shrink-wrap glare, each with its true item number, lot, quantity and case count. The project publishes how often a vision-language model reads each field correctly, and how many of its errors a second reviewer model flags for a person to check.
Why it is still open
Drone and robot vendors read labels from their own images, as Corvus does at ASICS, but none has published the images or an error tally. The closest public set, the Industrial Label Dataset released in June 2026, holds synthetic logistics labels and phone photos of printed copies, not labels photographed in racking.
Built with what exists
A vision-language model (see the Note "Three working devices"); a reviewer model (see the Note "Seven more flagged"); optionally a small model fine-tuned on one GPU (see the Note "One GPU to fine-tune"); a label printer; a phone on a telescopic pole, used from the floor.
First step
Print 100 labels, fix them to cartons on home shelving, photograph each at three angles under two light levels, and run a first tally.
Done looks like
1,000 photos with true values, released openly, with accuracy reported for each field under each condition (height, angle, light, damage) and the share of errors the reviewer model flagged.