I would think that with some wifi direction finders (I think I've read about such a thing) or maybe IR or other radio beacons in each corner of the store you could know pretty precisely where you are at any given point.
Also, I would imagine that some sort of RFID tag would be placed at each "destination" (IE, the Wheaties slot on the shelfs and the Diet Coke can slot). So robot could go to where he thinks it should be in general with it's load, and then find precisely (say, RFID tags go in the upper left-hand corner of a grocery slot).
The thing I can't get my head around is the imaging analysis. What I mean by that is what's been mentioned a few times above. One of these bad-boys has to:
- Recognize what's not supposed to be there and set it aside.
- Move everything that's supposed to be there into an orderly and presentable fashion.
- Add the new stock.
That probably requires it to scan the shelf and then do the recognition.
I would imagine setting this thing up (ie programming for specific tasks) might just take a specialized technician a few day at the site: take multiple pictures of each item for stocking (for image analysis I guess; and that's probably really standardized along with dimensions via the UPC and a barcode scanner), set out the RFID tags for each product with the positional reading, and observe a night of stocking. Oh and measurements for the dimensions of the shelf being targeted (IE, how much can we get into one place).
Hmm. Probably more complex than that. ;D