Key takeaways
- Each technology fails differently: RFID on tag cost, liquids and tampering; weight on look-alike masses and misplaced items; vision on occlusion and look-alike packaging.
- For weight sensing, load-cell resolution is rarely the limit. The natural weight variation of hand-made fresh food usually is.
- Vision plus weight works because their failures rarely overlap. Planogram design handles the cases where they do.
- When sensors disagree, hold the receipt and reconcile instead of guessing. Overcharges cost trust; undercharges cost margin.
- Count the full cost per SKU: tagging labor, weight profiling or image training, and dispute handling.
A smart cooler has about 10 seconds after the door closes to answer one question: what left the shelf? Everything downstream (the receipt, the inventory count, the restock order, the shrink report) depends on getting that answer right. There are three mainstream ways to get it: tag every item with RFID, weigh every shelf, or watch with cameras. Vendors like to argue about which is “most accurate,” which is the wrong question. Each technology fails on different items in different ways. The useful question is which failures you can live with, given your assortment.
What “knowing what you took” actually requires
A door session has to produce a receipt. To write it, the system must establish four things:
- Which items left the cabinet, by SKU.
- Which items came back, because shoppers pick things up, read the label and put them down again.
- Where things ended up, since a returned item on the wrong shelf will confuse the next session.
- That it all belongs to this session, not to a restock or a door left ajar.
There are two kinds of error, and they’re not equal. An overcharge (billing for something the shopper didn’t take) creates a dispute, a refund and a customer who now distrusts the device. An undercharge (missing something they did take) becomes shrink. Overcharges cost trust; undercharges cost margin. Your policy for ambiguous sessions should reflect that asymmetry.
RFID: tag every item
How it works. Each item carries a tag, typically passive UHF. Antennas inside the cabinet inventory every tag before and after the door session, and the difference between the two reads is the receipt.
Where it’s strong. Identity is exact. The tag says what the item is, so there’s no model to train and no confusion between two flavors that look alike. You can count the whole cabinet remotely at any time. RFID fridges have years of history in fresh-food vending; Byte’s smart fridges, for example, use RFID.
Where it struggles.
- Per-item cost. Every unit needs a tag, and someone has to apply it. Illustratively, if a tag plus application labor costs $0.10 and a cooler sells 1,500 items a month, that’s $150 a month per cooler, which can exceed the cost of the software that runs it. On a $2 drink, it’s 5% of the price.
- Physics. UHF signals and water or metal don’t mix well. Cans, foil pouches and dense liquids can detune tags or shadow them from antennas, so tag choice and placement need testing per package.
- Tampering. A removed, damaged or shielded tag is invisible. The cooler doesn’t know the item exists, so it can’t notice it leaving.
- Logistics. Untagged items can’t be sold, so every delivery depends on the tagging step happening upstream.
RFID suits programs with a central commissary that already labels every item, higher-priced fresh food and stable packaging.
Weight sensors: weigh the shelf
How it works. Load cells sit under each shelf or lane. When the door closes, the system compares each shelf’s weight before and after, then maps the change to the items planned for that shelf.
Where it’s strong. No per-item cost. It’s physics, so lighting and packaging redesigns don’t matter. It counts identical items precisely and detects restocks and anomalies (a door left open, a shelf leaned on) well.
Where it struggles.
- Look-alike masses. Two flavors of the same 12-ounce can weigh within a few grams of each other. On a single shelf, weight alone can’t tell them apart.
- Combinations. Say shelf 3 holds wraps at about 240 g, salads at about 300 g and yogurts at about 180 g. A 480 g drop is either two wraps or a salad plus a yogurt.
- Natural variation. This is the subtle one. Our load cells resolve ±2 g, but hand-made sandwiches and salads vary by far more than that from unit to unit. Sensor resolution is rarely the limit; product consistency is. Add ±10 g of natural variation to the example above and the two readings overlap almost completely.
- Misplaced items. A shopper takes a salad from shelf 2, changes their mind and puts it back on shelf 4. Weight reads a sale on shelf 2 and a mystery gain on shelf 4.
- Planogram discipline. Weight only works if restockers put items where the planogram says. One wrong lane breaks identity for that shelf until someone notices.
- Mechanical noise. Door slams, vibration and temperature drift need filtering and periodic re-zeroing.
Computer vision: watch the hand
How it works. Cameras inside the cabinet watch products move. Models running on the device classify each item by appearance and track whether it left the cooler or went back on a shelf, and where.
Where it’s strong. No per-item cost. It identifies by appearance, so two same-weight cans in different colors are easy. It sees where a returned item ended up. It can also spot empty facings and planogram drift, which helps merchandising.
Where it struggles.
- Occlusion. A hand wraps around a small item, or a shopper grabs two at once, and the cameras see only part of the action.
- Look-alike packaging. Flavor variants that differ by a thin color band, especially at an awkward angle.
- New and refreshed SKUs. The model needs reference images for every new item, and brands redesign packaging more often than you’d expect.
- Optics. Condensation on cold surfaces in humid rooms, glare and items hidden behind others all lower confidence.
Vision also carries a privacy question. Cameras should look at products, not people. Processing should stay on the device, and no images of faces should be stored or sent.
Side by side
| RFID | Weight | Vision | |
|---|---|---|---|
| Per-item cost | Tag plus labor on every unit | None | None |
| Identity comes from | Tag data | Weight on a planned shelf | Appearance |
| Same weight, different look | Strong | Weak | Strong |
| Same look, different weight | Strong | Strong | Moderate |
| Item returned to the wrong shelf | Strong | Weak | Strong |
| Hand-made fresh food | Strong if tagged | Weak (weight varies) | Moderate |
| Cans, foil, liquids | Needs testing | Strong | Strong |
| Tamper vector | Remove or shield the tag | Swap for an equal weight | Hide the item from view |
| New SKU setup | Tag supply | Weigh and assign a shelf | Reference images |
Why fusion works
Read the table down the weight and vision columns and a pattern appears: they fail in different places. A hand that hides an item from the cameras doesn’t change what the shelf weighs. Two same-weight cans that confuse the load cells look different to a camera. A salad returned to the wrong shelf is a puzzle for weight and an easy call for vision. When errors are mostly independent, combining the two catches far more than either catches alone.
An illustration, with example numbers rather than measurements: if vision alone were unsure about 1 session in 50, and weight independently resolved 9 in 10 of those, only about 1 session in 500 would remain ambiguous. The shape of that math is the point. On our own fleet, Kava Cooler’s vision plus weight fusion measures 99.4% checkout accuracy, audited against restock counts.
Fusion has one blind spot: items that look alike and weigh the same. Two flavors of the same sparkling water, side by side on one shelf, grabbed from a bad angle. The fix isn’t another sensor; it’s merchandising. Never put same-shape, same-weight variants next to each other on the same shelf. Put them on different shelves, and weight tells you which one moved.
When the sensors disagree
Every system eventually meets a session it can’t resolve. What it does next matters more than the headline accuracy number:
- Hold, don’t guess. Keep the receipt open for a short reconciliation instead of charging the most likely answer.
- Use the next restock. A restock count confirms what’s actually missing. Automatic restock reconciliation turns that into a routine check rather than an argument.
- Lean against overcharging. If it’s still ambiguous, charge the cheaper candidate or nothing, and book the difference as a known loss.
- Learn from it. Ambiguous sessions cluster on specific SKUs and shelves. Move those items. The same signals feed shrink analysis; our guide to where unattended retail shrink comes from shows how recognition errors look in your loss numbers.
The real cost per SKU
Sensor hardware is a one-time cost. The ongoing costs live in the life of each SKU. Before you choose, get answers to these on your actual assortment, not a demo shelf of soda:
- Onboarding: tag supply (RFID), a weight profile and shelf assignment (weight), or reference images (vision)?
- Restocking: tagging labor upstream (RFID), planogram discipline at the shelf (weight), or little extra (vision)?
- Packaging changes: nothing (RFID), only if the weight changes (weight), or new images (vision)?
- Disputes: what share of sessions needs review, and who handles them?
- Privacy: what is captured, where it’s processed and what is retained?
How 8Kava approaches this
We chose fusion. A Kava Cooler uses four edge cameras (six on the C2-XL) and per-shelf load cells with ±2 g resolution, with recognition running on the device’s own neural processor. Restock events are detected and reconciled automatically. Vision AI explains how the two signals are combined and how shrink detection uses them, and our Security & Trust page covers what the cameras capture: products, never stored images of faces.