Ashutosh Saxena’s Robot Learning Lab are the ones behind the ability for the PR2 to be able to figure out logical inferences concerning missing steps or poor instructions, and it uses its 3D camera in order to perform a scan of its environment, before getting a better idea on the kind of objects within.
For instance, it has been programmed to associate different objects with their capabilities, where a pan would mean one can pour stuff into it, and from it, while stoves will provide the allowance to place a corresponding utensil on them in order to heat stuff up later on. This means asking it to “heat water” would see it make use of the available tools, stove or microwave included. You do not even need to place the items in the same place, since the instructions will still work as it’s 3D camera does the work of scanning the kitchen.