Amazon Go

Why is this a machine learning problem? Why not use some low-power tag for items, and scan all items on customer exit?

Personal opinion: Amazon is not interested in creating supermarkets/wal-marts. Instead it wants to sell an ML solution to other brick & mortar stores, and this current effort is to prove plausibility. Selling an ML solution, with cameras and software, is harder to compete with than a tagging solution (especially if based on open and accessible hardware).

You're basically describing the old idea of 'tag everything with RFID' for the checkoutless grocery store. It didn't really catch on because it's a fairly expensive way to attack the problem given the small margins on most items and would require that EVERY manufacturer implement the tags or for the stores/distribution system to add tags to non compliant items which would also be fairly expensive.

So to get around needing to tag everything you can throw computer vision and machine learning at the problem to determine where the person is grabbing an item from and combined with well organized shelves you now know what they just grabbed off the shelves.