The part that really interested me was the process by which a robot orders a sandwich and pays for it. Unfortunately, the video fast-forwards through that bit.

The part that really interested me was the process by which a robot orders a sandwich and pays for it. Unfortunately, the video fast-forwards through that bit. The way the cashier looks at the cameraman suggests that there was some human assistance there. My point is that encoding probabilities about locations isn't enough. When the robot knew that the fridge was a probable place for a sandwich, it also knew how to open the door of the fridge, and that this was a required action to obtain a sandwich in that location. "Open the fridge door and look for a sandwich" is analogous to "Talk to the cashier and order a sandwich", but it seems like the robot can't do that yet. Subway also sells drinks, and if the person had asked for a drink, and subway had been found to be a probable location for a drink, the actions needed at the subway would need to be different to get a drink rather than a sandwich. It's cool research, and very useful, but it's worth pointing out that this robot can't yet actually fetch a sandwich from subway unassisted.