There isn't much published, but my hypothesis is:<p>Very accurate weight sensors on each shelf and a knowledge of which products are on each shelf should provide pretty accurate detection of which products a customer has picked up.<p>Cameras above shelves recognising simple things like t-shirt color detect <i>who</i> picked up that product.<p>That's probably good enough for a v1 prototype.<p>I guess now they combine that with per customer profiles (downrank products I'd never normally buy), and special weighing hardware where products are near identical (eg. Spice racks could have a different weight sensor on each column).<p>Combine all that with a human review system behind the scenes for resolving ambiguities.<p>Then remember that most people won't care too much if, out of a large grocery shop of 100 items, they were charged 50c for regular carrots when they picked up organic carrots for 65c.<p>Finally a simple dispute and refund system for cases which aren't caught by any of the above should keep customers happy.<p>Remember: Retail theft and shrinkage is about 1%. If you can make your system better than 1% while keeping customers happy, Amazon Go is good to go.