On-Shelf Availability Robots, OOS Detection Rate and Scan Cadence
At a glance: Out-of-stock detection is not one number. It is a chain of four measurements that each fail differently: how much of the shelf you actually photograph (capture rate), how reliably an empty facing is recognised (recall), how often a real product is called empty (precision loss, which produces false alarms), and how long the emptiness persists before a human acts (time to restock). A scanning robot that scores 95 percent recall but only photographs 60 percent of facings is a 57 percent detector, and that is the arithmetic that turns a pilot into a rejected purchase order.

Why "Out-of-Stock Detection" Is Not One Number
Buyers evaluating on-shelf availability robots ask a simple question, and they get an unhelpfully simple answer: "detection accuracy 95 percent". That figure usually refers to one link in a four-link chain, measured under studio conditions, on one planogram, at one camera height. When the same system is deployed in a 30,000 square metre hypermarket with six-metre gondola runs, reflective packaging and a night crew that restocks half the aisle mid-scan, performance separates badly by category and by aisle.
The chain has four links and they multiply, not average. Capture rate is the fraction of facings in the aisle the robot's cameras actually read at least once per scan, after occlusion by trolleys, other shoppers, end-cap displays and the robot's own turning circle are accounted for. Recall is the fraction of true empty facings the vision model flags. Precision is the fraction of flags that are genuinely empty, and its complement is the false-alarm rate that determines how much staff time the system wastes. Time to restock is the operational lag between flag and fill, which decides whether the alert was worth generating at all.
A system that reads 95 percent of facings, flags 92 percent of true empties and is 88 percent precise delivers roughly 0.95 times 0.92, or 87 percent of true out-of-stocks, with about 12 percent of its alerts being false. The same system on a gondola where occlusion cuts capture to 70 percent delivers only 64 percent. Neither number is wrong; the specification never said which one it was quoting. The rest of this guide is about writing a specification that says.

Capture Rate: The Number Nobody Specifies
Capture rate is the denominator of everything else, and it is the metric most often left out of vendor data sheets. It is measured as the count of distinct facings with at least one usable image, divided by the planogram facing count for the same aisle length. It is not the same as camera resolution or frames per second, and it is not improved by a better model.
Four physical conditions drive it. Camera geometry sets the vertical field of view, so a robot with a single forward camera pair looking along the aisle sees shelf edges at an oblique angle and misses the shelf back. Trolley and shopper occlusion removes facings transiently, and a single-pass scan converts that transient into a permanent gap for that cycle. Label placement on the shelf edge strip matters, because a missing or damaged label removes the anchor the vision system uses to locate the facing. And planogram drift, where the physical layout has moved away from the recorded plan, produces facings that exist but are not in the reference, which are typically discarded rather than counted.
| Condition | Effect on capture rate | What to specify |
|---|---|---|
| Oblique camera angle | Loses rear-shelf facings on deep shelves | Multi-pass scan with a second pass at the opposite direction |
| Trolley and shopper occlusion | 2 to 8 percent per single pass in trading hours | Two passes at least 20 minutes apart, or an overnight pass |
| Missing shelf-edge labels | Facings without an anchor get dropped | Report of unanchored facings, not silent discard |
| Planogram drift | Up to 10 percent of facings unaccounted | Drift report per aisle per week |
The practical consequence: a single nightly pass at 02:00 reads a clean, unoccluded shelf, which is the best capture rate the robot will ever achieve, and it is also the least useful, because by 02:00 the shelf is at its fullest and the out-of-stocks of the previous afternoon have already been noticed by customers. A mid-afternoon pass reads the shelf when it matters and captures less of it. The honest solution is a defined scan window combined with a stated capture rate for that window, tested in the aisles that matter, not in a demo aisle.

Recall, Precision and the Cost of a False Alarm
Vision models for shelf scanning are typically tuned on a trade-off between recall and precision, and the two pull in opposite directions. Raising the threshold to catch every empty facing means flagging shadowed gaps, wrapped promotional displays, facings temporarily removed for a pick, and SKUs whose packaging is itself dark and low-contrast. Lowering it means missing real holes.
The right setting is a commercial question, not a machine-learning question, because the two error types cost different things. A missed out-of-stock costs lost sales on that SKU until a human notices by walking the aisle, which for a fast-moving consumer goods line in a high-traffic store can be several hundred units of foregone revenue per day. A false alarm costs a few minutes of a colleague's time to walk over and confirm, plus a slow erosion of trust in the alert stream, which is the more expensive of the two because once the team stops believing the alerts the whole system becomes shelf theatre.
| Error type | Operational cost | Typical acceptable rate |
|---|---|---|
| Missed out-of-stock (recall loss) | Lost sales plus a customer who finds the gap | Under 10 percent on high-velocity SKUs |
| False alarm (precision loss) | Staff walk time plus loss of trust in alerts | Under 15 percent per aisle per cycle |
| Wrong SKU identified | Restock sends the wrong product, double cost | Under 3 percent |
| Duplicate flag for one hole | Alert fatigue, no direct cost | Deduplicate within a 30-minute window |
Two design choices move these numbers more than model tuning. First, confirmation by a second pass: a hole flagged on pass one and still empty on pass two twenty minutes later is almost certainly real, and that single rule removes most transient false alarms from trolleys and pickers. Second, class-aware thresholds: applying a stricter threshold to a high-velocity beverage facing than to a slow seasonal line matches the cost of error to the value of the SKU. A vendor that cannot express thresholds per category has not built a retail product, only a detector.

Scan Cadence Arithmetic: How Often Is Often Enough?
Cadence is where most business cases are won or lost, because it multiplies directly against fleet size and against battery and dock capacity. The arithmetic is straightforward once you accept a target: the maximum time a hole should be allowed to persist before the system notices it.
Start from the aisle length and the robot's effective scanning speed in real trading conditions, not on an empty floor. A wide-aisle scanning robot covering both sides of a gondola run at roughly walking pace with stops completes around 1,200 to 1,800 linear metres of shelf per hour when shoppers are present, and closer to 2,500 metres per hour overnight. A 25,000 square metre store with 4,000 linear metres of shelf therefore needs about two and a half to three hours per full pass in trading hours.
That single fact determines the practical cadence. If a full pass takes 2.7 hours, then "scan every hour" is not a cadence the store can be given by one robot; it is a statement that some aisles are scanned every hour and others every three, or that three robots are required. The specification should therefore name a maximum gap per department rather than an average cadence for the store, because the fresh department with high velocity and high shrink exposure deserves a tighter loop than the seasonal aisle.
| Department | Velocity | Target max gap | Passes per 12 trading hours |
|---|---|---|---|
| Fresh and chilled | High | 90 minutes | 8 |
| Beverages and household | High | 2 hours | 6 |
| Grocery ambient | Medium | 3 hours | 4 |
| Seasonal and general merchandise | Low | Once daily | 1 |
Battery and dock capacity follow from that table rather than preceding it. A weighted average of four to five passes across a 2.7-hour full-store pass is 11 to 13 hours of scanning per trading day, which is more than one eight-hour battery shift can deliver, so the fleet size is set by shift coverage, and the dock scheduling and charge window planning becomes the constraint that decides whether the fourth pass actually happens. This is exactly the failure mode in scanning pilots that report good accuracy and disappointing business results: the accuracy was real, the cadence was aspirational.

What the Alert Actually Has to Contain
An alert that says "out of stock in aisle 14" is not actionable; a colleague still has to walk the aisle and work out which facing. The alert has to carry enough structure that the recipient can act without a second visit. In practice that means five fields, and the absence of any one of them is the most common reason a scanning system is quietly abandoned after three months.
- Aisle, bay and shelf level, so the colleague walks to a shelf, not a department.
- SKU identifier mapped to the store's own item file, not the vendor's catalogue number, because the restocker works in the retailer's system.
- An image of the facing, so the colleague can confirm from the aisle entrance and skip the walk if the image shows a trolley.
- A confidence figure and the reason for the flag, so a repeated low-confidence flag on a reflective display is understood as a known blind spot rather than an empty shelf.
- Time of first flag and elapsed time, which is the only field that lets a store manager measure whether the loop closes.
The last field is the one that converts the system from a detector into a management instrument. Once each alert carries elapsed time, the store can report a time-to-restock figure by department, and that figure drives staffing decisions far more directly than a headcount reduction case. A pilot that reports "we found 240 out-of-stocks" has told management nothing durable; a pilot that reports "median time from first flag to fill fell from 5.5 hours to 2.1 hours across chilled" has told management something they will fund.
Evaluating a Pilot Without Being Misled
Pilots are usually designed to demonstrate that the technology works, which is the wrong test. The test that matters is whether the loop closes and whether the numbers survive a hostile re-measurement. Four checks separate a real trial from a demonstration.
First, ground-truth a sample by hand. Pick 20 aisles at random, have a colleague walk them with a printed planogram while the robot scans, and count the real holes. That gives an independent recall and precision figure for the same shift, and the comparison against the vendor's report is the single most informative number in the pilot. Second, count the alerts that nobody acted on. Alert volume is vanity; acted-on alerts are the product. Third, measure capture rate by aisle, not store-wide, because a store average hides the aisles with 40 percent capture behind the ones with 98 percent. Fourth, run one week in trading hours and one week overnight and compare, because the overnight numbers will flatter the system and the trading-hours numbers are the ones the store actually lives with.
There is a cross-check available in the store's own data that costs nothing: compare the detected out-of-stocks against the electronic point-of-sale gaps for the same SKUs and the same days. A shelf hole that was never sold out at the till is either a false alarm or a stocking error, and both are worth knowing. Where the store has already invested in electronic shelf labels, the comparison of scanning robots against electronic shelf labels sets out which of the two instruments actually produces the availability number, and the planogram compliance workflow covers the related question of whether the shelf looks right rather than merely whether it is full.
Specification Clause Set for Availability Scans
Written as purchase-order clauses rather than aspirations, the following five lines turn a vague capability claim into something that can be tested at acceptance. Each clause names a metric, a measurement method and a failure consequence, which is the difference between a specification and a wish.
| Clause | Metric to specify | How acceptance is tested |
|---|---|---|
| Capture | Capture rate by aisle, in trading hours, minimum 90 percent on the top three shelf levels | Manual facing count against robot facing count on 20 random aisles |
| Detection | Recall minimum 90 percent and precision minimum 85 percent on a seeded test set | Known holes created before a scan, then counted in the alert export |
| Cadence | Maximum gap per department as tabled above, measured over five consecutive trading days | Timestamped scan log per aisle, no averaging across departments |
| Alert content | The five fields listed above, exported in a documented schema | One week of raw alert export delivered to the buyer, not a dashboard screenshot |
| Loop closure | Time-to-restock reportable per department from the system itself | Report produced at day 30 and day 90 of the pilot |
The data export clause is worth insisting on for its own sake. A scanning robot generates a per-aisle, per-hour record of shelf state, and that record has value beyond restocking: it is the evidence base for promotional compliance, for shelf-space negotiations with suppliers, and for the retail-media inventory question of how many shoppers actually passed a display. A scanner that only exposes a dashboard the retailer does not own has handed the retailer a service, not an asset. The retail media business case depends on exactly this kind of first-party measurement, and the on-shelf availability overview covers the wider workflow the scanner sits inside.
None of this requires a new class of machine. The scanning payload is a camera mast and a compute module, and the vehicle underneath it is an ordinary indoor autonomous platform chosen for aisle width and battery endurance, of the kind covered in the deployment throughput planning guide. Getting the arithmetic right before the order is placed is what makes the difference between a scanning programme that reports a number management trusts and one that reports a number nobody checks.
