Examples: https://www.youtube.com/watch?v=p1LykdRWTfk https://www.youtube.com/watch?v=-hCA8L7v-R8
What is really needed is a way to talk to multiple RFID tags quickly without any crosstalk. Think of a handbag with a dozen candy bars in it. How do you scan them all quickly without missing any?
The goal has always been to scan something large in one pass, say a shrink-wrapped pallet of items that could number in the hundreds or thousands. Obviously this is technology that Amazon could benefit from as well, I wouldn't be surprised if there's a lot of technology overlap here.
Most RAIN Rfid readers have a read rate of 800+ tags/second. Door portals are a solved solution and now phased arrays are on the market which scan all items in realtime. Take a look at the impinj marketing material for more information. For example the RS2000 chip: http://www.impinj.com/products/reader-chips/indy-rs2000/
But the video clearly shows the items being recognized as someone takes them from a shelf (and puts them back). The items don't need to be recognized at the time you walk out, the store just needs to know that you've walked out.
I think that's just visually giving you an idea of what's happening. Like I said earlier, I'm giving a simple naive presentation of how the system might work. Or, at the least, how I would design it without dealing with finicky shelf sensors. Ask anyone that has ever worked in a hotel with in-room minibars how well those things pan out in real life.
Unless that store is just a mockup and not how it really looks and works, I see nothing on those shelves or in the sky above it that is watching you put that cupcake back on the shelf.
The real way to find out is if someone in the demo video can pull their phone out of their pocket mid-trip and see their current inventory and total price. But I didn't see that in the video.
The girl in the video at 1:26 looks at her list and total, but after she's been through the exit turnstyle.
The video makes it look like it's constantly up to date, but that sounds quite complicated (despite dropping the ML, DL buzz words).
A friend at Amazon said he tried to trick it by taking two items at a time and it got it right every time, so yeah it does appear to track when you take the items.
From my first quick take of the video, the app+turnstyle is used to identify you to the store. The video system then tracks your position as you walk around.
When you walk out, the items are recognized and tallied by a large RFID sweep. Funneling you back out through a turnstyle makes sure the vision system knows it's you. Notice that you don't need to barcode yourself on the way out, and the exit system is phone agnostic (it's not checking for an NFC or Apple Pay tag or anything).
The whole "tracking individual items as they come on and off the shelves" task is a very complex thing. But tracking bodies as they walk around a 1,500 square foot room isn't that hard.