← AI in physical storesC2 · LESSON 2 OF 7

C2, Lesson 2: Shelf Intelligence: Robots, Cameras and the Empty Slot

Tags: Sourced / General / Our view / Illustrative. Reading time about 13 minutes.

1. The story: the robots that went home

In 2017 Walmart began rolling out six-foot-tall robots that rolled the aisles with cameras, scanning shelves for empty slots and misplaced products. The robots came from Bossa Nova Robotics of Pittsburgh. [Sourced: Reuters, Nov 2-3, 2020; WSJ, Nov 2, 2020; AP] Walmart had made them a frequent topic at media and investor events, hoping they would cut labor cost and raise sales by keeping products in stock. [Sourced: WSJ, Nov 2, 2020] On November 2, 2020 Walmart said it had ended the partnership, after about five years and roughly 500 stores. In its statement it said "this was one idea we tried in roughly 500 stores just as we are trying other ideas in additional stores." [Sourced: Reuters, Nov 2-3, 2020] The Wall Street Journal, which reported it first and cited people familiar with the matter, wrote that during the pandemic Walmart found human workers, who had been walking the aisles to pick online orders, could provide similar information, and that there was some concern about how shoppers reacted to the robots. [Sourced: WSJ and AP, Nov 2, 2020; unnamed sources, not Walmart's own explanation] The lesson: the robot could see a gap. Seeing the gap was not the scarce thing. People walking the store for other reasons could see it too. [Our view]

2. The one idea

An empty slot costs money only until someone fills it. Shelf intelligence shortens the time between "the shelf is empty" and "the shelf is full." It works when detection is the slowest step, and fails to pay when the slow steps are elsewhere: ordering, the backroom, the truck. [General]

3. The kitchen-table version

You run out of bottled water. Maybe the shelf is empty but there is a case in the backroom. Maybe you never ordered enough. Maybe the truck is late. A camera telling you "the shelf is empty" helps with only the first. If most of your empty shelves come from the second and third, a camera is an expensive way to learn what you already knew. [General]

4. How big is the out-of-stock problem?

A widely cited study, "Retail Out-of-Stocks: A Worldwide Examination of Extent, Causes and Consumer Responses" by Gruen, Corsten and Bharadwaj (2002, for the Grocery Manufacturers of America), estimated the average out-of-stock rate worldwide at 8.3 percent of items, and above 10 percent for faster-selling and promoted products. [Sourced: GMA 2002 study; old data, mostly 1990s grocery] What happens when a shopper meets an empty slot: on average 31 percent go to another store, 26 percent substitute another brand, and 9 percent do not buy; the release put the average sales loss at 4 percent. [Sourced: FMI release, Oct 14, 2002 (averages, verifier-confirmed); the study synthesizes 52 earlier reports and the shopper survey covers more than 71,000 shoppers] Why shelves are empty: about half of out-of-stocks come from store ordering and forecasting, and about a quarter from poor shelf management (the product is somewhere in the building but not on the shelf); a further 28 percent arises upstream (distribution center, retail head office or manufacturer, other). The study's own exhibit gives 47% store ordering and forecasting, 25% shelf management ("in the store, not on the shelf") and 28% upstream, and the FMI release says the same. [Sourced: Gruen, Corsten and Bharadwaj, GMA 2002, "OOS Causes: Worldwide Averages" exhibit and Exhibit 23; FMI release, Oct 14, 2002; both read Oct 7, 2026] These numbers are over twenty years old, so use them as a method, not a measurement of any store today. [Our view]

5. Four ways to see a shelf

MethodHow it worksGood atLimits
Human walkAssociates check and countFlexible, can fix on the spotSlow, inconsistent
Roving robotWheeled camera scans aisles on a scheduleSame route, repeatableCost, navigation, shopper reaction
Fixed camerasCameras on ceilings or shelvesContinuousInstall cost, privacy, coverage
Shelf sensors or RFIDWeight sensors or tagsExact countsTagging cost, setup
[General]
All of them make an alert. Value comes only from what happens after (Lesson 1).

6. Worked example: what is the scanner worth?

All numbers [Illustrative] except the study percentages cited above; computed by script. A store has 12,000 items on shelf. At the study's 8.3% average, about 996 slots are empty at any moment. Suppose each empty slot would sell $4 a day. That is 996 x $4 x 365 = $1,454,160 a year of sales behind empty slots. Use the release's average reactions: 31% go to another store and 9% do not buy, so about 40% of that is lost to your store (the 26% who switch brands still buy from you): $581,664 in lost sales. At 28% gross margin: $162,866 a year. Now split by cause, using the 47% / 25% shares:

  • Fix ordering and forecasting (47%): up to $76,547 a year.
  • Fix shelf management (25%): up to $40,716 a year. That is the only part a shelf scanner can speed up, and suppose it catches half: $20,358. A robot or camera system that costs $50,000 a year to run would not pay here. Better forecasting (Commerce Module 4) or replenishment rules (Commerce Module 9A) could be worth more, for less. Change the inputs (more sales per slot, a much higher empty rate, cheap hardware) and the answer can flip. That is the point: do the arithmetic for your store before buying. [Our view] This also fits what Walmart's own words and the WSJ reporting suggest: there was no cheap way to prove the robots beat people already in the aisles. We do not know Walmart's internal numbers. [Our view; no claim about Walmart's data]

7. What it costs and what it breaks

  • Cost: hardware, install, maintenance, connectivity, plus associate time on the alerts.
  • False alerts: shelf look-alikes (a package in two sizes) and wrong-place stock produce wrong alerts.
  • Shoppers: the WSJ report mentioned concern over how shoppers reacted to robots. People notice.
  • Privacy: cameras in aisles see people. Say what is collected and keep it to shelves. [Sourced: AP, Lesson 1]
  • Wrong target: fixing detection when the cause is ordering. The ordering part needs forecasting, pack-size and lead-time fixes.

8. What retailers say they do now

Walmart says it continues to test technologies and to invest in its own processes and apps to track inventory and move products to shelves faster. [Sourced: Walmart statement via Reuters, Nov 2020] In March 2026 Walmart said about 2,300 US locations use digital shelf labels (Lesson 5). [Sourced: Walmart corporate post, Mar 2, 2026; company claim]

9. Monday checklist

  1. Walk 20 fast-moving items twice a day for a week. Count empty slots and whether product was in the backroom.
  2. Sort empties into cause: ordering, backroom, late truck, theft or damage.
  3. Value each cause with the Section 6 method.
  4. Only if "on the shelf somewhere else" is the biggest cause, price a scanner.
  5. Decide who acts on an alert and how fast.

10. Try it yourself (no code, 30 minutes)

Redo Section 6 with your numbers: SKUs, empty rate (measure it), sales per slot per day, margin. Compute the value of fixing each cause. What is the most you could pay a year for a tool that fixes only shelf management?

11. Self-check

  1. Why did Walmart end the Bossa Nova robots in November 2020? Walmart's statement gave no detailed reason; the WSJ, citing people familiar, said humans could get similar results and there was shopper-reaction concern.
  2. What share of out-of-stocks came from store ordering and forecasting in the 2002 study? About 47%, roughly half.
  3. When a shopper faces an empty slot, how many go to another store (average)? About 31%.
  4. A scanner only helps with which cause? Shelf management: product exists in the store but is not on the shelf.
  5. In our example, why does a $50,000 scanner not pay? It can only fix about a quarter of $162,866 of margin, and catches half of that: $20,358.

12. Where this is going

Shopping agents will ask "is it on the shelf nearby?" Stores that can answer accurately will be the ones the agent picks (C1, Lesson 7). A shelf count that is right matters more when software asks. [Our view]

Sources

Review flags

  • The WSJ explanation relies on unnamed sources; Walmart itself did not give those reasons. Lesson says so.
  • RESOLVED with limitation (Oct 7, 2026): the WSJ page is paywalled beyond its headline, standfirst and lead ("after finding during the coronavirus pandemic that humans can help get similar results"). AP (Nov 2, 2020) reports WSJ "citing unnamed people familiar with the situation who said the retailer found human workers could get similar results. There was also some concern about how shoppers reacted... according to the report." CNBC (Nov 2, 2020) adds that the report said Walmart U.S. CEO John Furner worried about shopper reactions; the lesson does not state that. The lesson attributes both reasons to WSJ's unnamed sources, not to Walmart, and quotes no more of the WSJ than the lead. No change to the lesson text was needed.
  • RESOLVED (Oct 7, 2026): the 47% / 25% split. The full GMA study text (read) shows 47% store ordering and forecasting, 25% store shelving and 28% upstream in its causes exhibit (Exhibit 23 splits ordering 34% and forecasting 13%, shelving 25%, distribution center 10%, retail HQ or manufacturer 14%, other 4%). The executive summary says "about one-half," "about one-fourth" and "about one-fourth" in words; the exhibit numbers are the exact ones. Also confirmed in the study: 8.3% average out-of-stock rate, 52 studies, more than 71,000 consumers.
  • The figures are from 2002 and mostly grocery; the lesson labels them as such.
  • The $4 per slot per day, 28% margin and $50,000 tool cost are illustrative.

Claim ledger (C2 Lesson 2)

#ClaimSourceVerifiedStatusRe-check
1Walmart began rolling out shelf-scanning robots in 2017Reuters2026-10-06Readnone
2Ended Bossa Nova Nov 2, 2020; roughly 500 stores; five yearsReuters, AP2026-10-06Readnone
3Walmart quote "one idea we tried in roughly 500 stores"Reuters2026-10-06Readnone
4WSJ (via AP): unnamed people familiar said humans could get similar results; some concern about shopper reactionWSJ headline/standfirst and lead (paywalled beyond); AP, Nov 2, 2020 (read); CNBC, Nov 2, 20202026-10-07Resolved with limitation: attributed in the lesson to WSJ's unnamed sources, not to Walmart; WSJ body not read in fullnone
5Average OOS 8.3% worldwide; over 10% for fast/promotedGMA 20022026-10-06Readnone
6Averages: 31% another store, 26% substitute, 9% no purchase; 4% sales loss; 52 earlier reports synthesizedFMI release 20022026-10-06Verifier-confirmednone
747% ordering/forecasting; 25% shelf managementFMI release 20022026-10-06Read excerptconfirm in full study
8Worked-example arithmeticOur computation2026-10-06Illustrative, script-checkednone
7Causes: 47% ordering and forecasting, 25% shelf, 28% upstreamGMA 2002 study exhibits; FMI release2026-10-07Re-read in full studynone