How an AI trolley keeps track when you change your mind

Imagine putting a jar into a trolley, then deciding against it and returning it to the shelf. A self-scanning shopper must reverse an entry. A trolley that recognises groceries for itself has to work out what changed. Its running bill is a record of decisions made during the shop, not just a list of objects it has seen.

Morrisons and Instacart launched their Preston trial on 28 July 2026. The Caper trolley recognises items as they enter, weighs fresh produce and displays a running basket total. Shoppers can bag their purchases as they go and scan a Morrisons More Card for loyalty offers. At this store, payment finishes by scanning the trolley’s displayed barcode at a dedicated Caper checkout lane.[1]

A camera and a scale answer different questions

Instacart’s engineering account describes basket-facing cameras that help locate items in three dimensions. Several objects may move together; hands, bags and other groceries can obscure a label. Weight supplies another clue when the visual view is incomplete. But weighing has its own complication: moving the trolley disturbs the reading, which can take one to three seconds to settle after a change.[2]

The system combines processing on the cart’s NVIDIA Jetson computer with models interpreting visual context in the cloud. Outputs from those paths combine to interpret shopping activity. The architecture described by NVIDIA therefore shares work between trolley and remote computers; the presence of onboard hardware does not mean every part runs locally.[3]

For the bill, the sequence matters. Two identical jars added together should create two purchases. Lifting one jar to make room for something else should not create a third. This is why a useful shopping system has to interpret changes across time, as well as recognise a product in an image.

Putting intelligence into the trolley

Instacart markets Caper as hardware that works with existing checkout, payment and loyalty systems. Cameras, sensors and weighing equipment sit in the cart, and its product documentation says deployment does not require store modifications. That describes the general platform; each retailer’s checkout arrangement still matters, as Preston’s final barcode payment demonstrates.[4][1]

For a grocer, this offers a way to introduce the service while retaining familiar systems around it. For the shopper, the running total moves an important piece of information earlier in the trip: a spending decision can be made beside the shelf, before arriving at payment.

The screen has another customer

The commercial proposition also includes selling screen space. Instacart’s advertiser documentation describes campaigns appearing when a trolley enters an appropriate aisle and the promoted product is in stock. Eligible advertisers compete in an auction; one second of display counts as an impression. That service applies at advertising-enabled locations. The documentation does not establish that these advertising arrangements are active in the Morrisons trial.[5]

A displayed basket total and a sponsored suggestion can therefore serve different purposes on the same device. One helps a shopper understand an existing commitment; the other invites a new purchase. A relevant discount may be useful, but it is still worth distinguishing a lower price on something wanted from encouragement to add something unplanned.

The small act of putting something back is a useful test of the experience. The screen should follow that decision cleanly, leaving the shopper free to change their mind without reconstructing the bill.

Report an error or suggest a correction