Shelf Scanning Robots vs Electronic Shelf Labels: Which Job Each Solves

Four Jobs That Get Confused

Electronic shelf labels and inventory robots are frequently evaluated as alternative answers to the same question. They are not. The store has four distinct jobs, and each technology reaches only some of them.

Job	What it means operationally	ESL reaches it	Robot reaches it
Price display	The shelf shows the correct price to the shopper	Yes, directly	No, it only observes
Price integrity audit	Displayed price matches point of sale for every facing	Yes, if POS is the single source	Yes, independently verified
On-shelf availability	A product is where the planogram says it should be	No	Yes, the core use case
Planogram compliance	Facings, facings per SKU, and position match the layout	Partly, position only	Yes, position and presence
Recall and date-code sweep	Specific batches are located and pulled	Only if barcode-coded	Yes, by scan

The confusion arises because an ESL deployment genuinely eliminates one class of problem, the mis-priced shelf, and the vendor's case study then implies the store no longer needs to audit price at all. That holds only while the POS-to-label link is intact. Where labels are updated by batch overnight while price changes are authorised during the day, the ESL and the till still disagree for several hours, and nothing detects it unless something reads the shelf.

The inventory job is a separate matter. An ESL has no idea whether the facing behind it is stocked, and no ESL deployment in service today tells a store manager that a promotion is eighty percent sold out at 2pm on a Saturday.

Accuracy, Task by Task

Published accuracy claims are not comparable across technologies because they measure different things. The table below uses the definitions a store operations team would apply, with the range observed across deployments rather than a single vendor figure.

Measurement	ESL network	Shelf-scanning robot	Manual audit
Price displayed correctly	99.4 to 99.9 percent while in sync	Detects the exceptions, does not fix them	95 to 98 percent
Detection of a wrong price	None if the link is the source of truth	94 to 98 percent per pass	Depends on sampling, typically 40 to 60 percent
On-shelf availability detection	Not applicable	90 to 96 percent, lower in crowded categories	60 to 75 percent, and only where staff look
Planogram position compliance	Label position only	85 to 93 percent at SKU level	Highly variable
Audit frequency achievable	Continuous for price	2 to 5 passes per store per week	1 per quarter at most
Cost per store per year	Label capex plus gateway and battery	Subscription or capital plus service	Store labour hours

Two readings of this table matter. First, the robot is not more accurate than an ESL at knowing the right price, because the ESL does not need to measure anything; it displays whatever the system tells it. Second, the robot is the only one of the three that produces an independent, dated record of what was on the shelf, and that record is what makes the availability and compliance jobs measurable at all.

Accuracy also degrades differently. ESL accuracy depends on link integrity and label battery health, both of which fail quietly at scale. Robot accuracy depends on lighting, shelf density and occlusion, and it degrades in specific known conditions such as deep promotional bins and end caps with stacked product.

The Capital and Subscription Arithmetic

The comparison only becomes a decision when it is run on a concrete network. The model below assumes 25 stores, an average of 4,500 SKUs per store, and a mix of 18 large-format and 7 small-format sites. All figures are planning estimates and should be replaced with quoted prices before approval.

Line	ESL network, 25 stores	Shelf-scan robot programme, 25 stores
Hardware	112,500 labels at USD 6.50 = 731,250	18 large stores at 1 unit, 7 small at 1 shared unit = 25 units
Unit price	Included above	USD 21,000 per unit including dock and software
Network and gateway	USD 2,800 per store = 70,000	Wi-Fi survey and AP works, USD 1,200 per store = 30,000
Installation labour	USD 4,200 per store = 105,000	Commissioning included in unit price
Year 1 total	906,250	555,000
Annual running, years 2 to 5	Battery and gateway support, USD 1,900 per store = 47,500	Service at 9 percent of hardware = 47,250
Five-year total	1,096,250	744,000
Annual cost per store	8,770	5,952

On capital alone the robot programme is the cheaper of the two across a 25-store network, which is not the answer most operators expect. The reason is that ESL cost scales with the SKU count, roughly 112,500 labels in this model, while robot cost scales with store count. A network with 8,000 SKUs per store reverses the ranking: ESL cost roughly doubles while the robot fleet is unchanged.

That single relationship is the most useful output of this exercise. ESL economics are driven by SKU density; robot economics are driven by store count and floorspace. A convenience format with 2,400 SKUs and a small footprint should lean on labels. A hypermarket with 30,000 SKUs should not treat labels as the answer to availability at all.

What the Robot Adds That Labels Cannot

The availability job has a revenue mechanism that is worth stating precisely, because it is usually quoted as a percentage range without arithmetic. The standard framing is that one to four percent of sales are lost to items that are out of stock but still listed as in stock, and that shoppers substitute or leave.

Worked on a single large store doing USD 18 million a year, a two percent recoverable share is USD 360,000 of sales. At a 26 percent gross margin that is USD 93,600 of gross profit, against an annual robot cost of about USD 7,000 at the figures above. The ratio is large enough that even a tenth of the assumption pays for the programme, which is why the modelling discipline matters more than the headline: the recoverable share must be the share the store can actually close, meaning the shelf gap is fillable from the back room within the same trading day.

Availability outcome	Typical recoverable share	Mechanism
Gap with stock in the back room	50 to 70 percent of detected gaps	Task generated to the right aisle, filled same day
Gap with no stock on site	0 percent from the robot alone	Feeds replenishment forecasting, not recovery
Wrong facing, product present	80 to 90 percent	Corrected on the next pass
Promotion sold out early	Depends on supply	Detected hours earlier than a manual walk
Recall still on shelf	100 percent of detected units	Compliance exposure, not revenue

There is also an audit trail argument that matters to larger networks. Because the robot produces a dated per-facing record, the availability figure in the monthly operations pack stops being an estimate from a sample walk and becomes a measured census. Retailers who have run both report that the biggest change is not the detection rate but the end of arguments about whether a store was in compliance on a given day.

Running Both Without Double Paying

The sensible end state for most networks is both technologies, with a clear division of labour and an explicit rule about what the robot is not asked to do.

Responsibility	Assigned to	Rationale
Price shown to shopper	ESL	It is a display device; use it as one
Price-to-POS validation	Robot, sampled weekly	Independent check on link integrity
Availability detection	Robot	Labels cannot see stock
Replenishment task generation	Robot software into task system	Avoid a second work queue
Date-code sweep	Robot, on a scheduled pass	Consistent coverage, dated record
Promotional end-cap audit	Robot where reachable, manual otherwise	Camera limits on deep bins

Three integration points do most of the work. The robot's task output should land in the same queue the store already uses for replenishment, or the store will run two lists and complete neither. The price exception list should route to whoever owns the price file, not to the store manager, because the store cannot fix a POS error. And the availability data should feed the replenishment forecast, not just the daily task list, or the same SKUs will show as gaps every week.

The double-payment trap is subtler than duplicate hardware. It is the store that buys labels to solve price, then buys a robot to solve availability, then asks the robot to do a price audit that the labels already guarantee, and finally pays for a nightly manual walk because nobody trusts either system's data. The fix is procedural: one owner per job, one task queue, and a stated audit frequency for each.

Rolling Out Across 25 Stores

Networks that deploy the whole estate at once tend to discover the integration problems twenty-five times simultaneously. A staged rollout over three quarters costs the same in hardware but surfaces the process issues on one store.

Phase	Stores	What is being proven	Exit criterion
Pilot	2, one large one small	Route coverage, dock siting, Wi-Fi	Full-store map in under 6 hours
Measurement	Same 2, 90 days	Detection accuracy against manual audit	Within 5 points of vendor claim
Integration	Same 2	Task queue and forecast feed	Tasks closed same day above 70 percent
Wave 1	6	Operational load on store teams	No additional hours rostered
Wave 2	10	Service and spares logistics	Under 2 percent downtime
Wave 3	7	Steady state	Fleet dashboard covers all sites

Two practical constraints shape the schedule. Store mapping takes a supervised first pass of three to six hours per store, and stores with seasonal layout changes will need re-mapping when the layout changes materially. Dock siting usually lands in the back-of-store corridor, and the constraint is power rather than space. Both are worth resolving in the pilot rather than in wave three.

For networks where the same estate also runs cleaning robots, the survey and network works can be shared, and the multisite deployment sequencing published by AOMAN sets out how to sequence the two programmes through the same store calendar. Stores that are formats of the same banner can also reuse planogram definitions across sites, which shortens mapping considerably.

When the Robot Is the Wrong Purchase

The case above assumes a store with a stable planogram, workable aisle widths and a task system that can absorb new work. Four conditions make the shelf-scanning robot a poor investment, and they are worth checking before any pilot is approved.

• High SKU churn with weekly layout resets. Re-mapping cost can exceed the availability gain, and the record goes stale between passes. Retailers with frequent promotional resets should either map against the base planogram only or wait.
• Narrow aisles with dense promotional clutter. Detection rates fall sharply where product overhangs the shelf edge and where pallet displays block the aisle. Measure the detection rate in the actual cluster before committing.
• No owner for the task output. If nobody closes the replenishment tasks the robot generates, the store pays for detection and gets no recovery. The programme must have an operations owner, not just a technology sponsor.
• Very low SKU density. Below roughly 3,000 SKUs per store with high staff presence, manual walks plus price labelling may be adequate and the robot's value is primarily the audit record, which is a compliance purchase rather than a revenue one.

The mirror case is just as important. A label-only strategy stops at price. If the store's actual loss is driven by availability, labels deliver a clean price file and no improvement in the shelf gap, and the operations team will spend the next two years manually auditing a problem the label network was never designed to see.

Frequently Asked Questions

Do I still need shelf-scanning robots if I have electronic shelf labels?

Yes, if on-shelf availability or planogram compliance matters to your sales. Electronic shelf labels show the price the system sends them; they cannot detect that a facing is empty, misplaced or occupied by the wrong SKU. The robot covers the shelf-state jobs, the labels cover the display job.

How accurate are inventory robots in large-format grocery stores?

Between 90 and 96 percent detection on on-shelf availability, and 85 to 93 percent on planogram position at SKU level, with lower rates in deep promotional bins and where product overhangs the shelf edge. Accuracy is best validated by running the robot and a manual audit on the same aisles for two weeks.

Which is cheaper across a 25-store network, labels or robots?

It depends on SKU density more than store count. At 4,500 SKUs and 25 stores, labels cost about USD 1.1 million over five years and a single-unit-per-store robot programme about USD 750,000. At 8,000 SKUs per store the ranking reverses, because label cost scales with SKUs while robot cost scales with stores.

What is a realistic out-of-stock recovery rate?

Between 50 and 70 percent of detected shelf gaps are fillable from back-room stock during the same trading day. Gaps with no on-site stock are not recovered by the robot; they feed replenishment forecasting instead.

The platform used in the arithmetic above is described on the AOMAN D1 delivery robot product page, which also lists the sensor set that determines detection accuracy at shelf height.