Few stores have the exact same layout for years on end, which would make a robot economically viable.
That's not really a problem: any robot sophisticated enough to load shelves will have enough processor power to carry re-loadable maps of the layout.
I think the biggest challenges are making the gripper flexible enough to pick up a huge variety of packages, and making the vision system sophisticated enough to recognize when the shelves are "ordered" and "fully stocked".
The gripper(s) would need to perform tasks like lifting a heavy can of juice up and out of a shipping case, then changing its grip so it could slide the can horizontally onto the shelf.
The vision system would have to recognize whether a customer had discovered a big sale on root beer and abandoned his Coke there instead of returning it to where he got it. It would need to determine whether there's a big hollow space behind the cans neatly arrayed across the front of a shelf. And deal with a lot of other situations that are easy for specially-evolved biological systems, but very hard for designed-from-scratch mechanical ones. I'm sure there are projects in various labs that can handle such challenges. Possibly even a few systems being tried out in the field. But I'm guessing they're still too spendy to deploy in a supermarket.