I actually work in this field and it's a lot more complicated than it sounds. When you're training AI to recognize products in a store, you have a set list of products it needs to be trained on. A person might go to many different stores which increases the possible variation of products exponentially. Amazon's model is also much more complex than just cameras, involving weight sensors in shelving, pressure detection, facial recognition. A store where everything is laid out in predictable, well lit, organized rows is already a nightmare. A fridge, even if it's way smaller, is way, way less predictable