I don't know whether this is impressive or not. It's not like they took three robots, told them the puzzle and they started thinking and talking. It depends on how they actually implemented it. With prolog it would take 30 minutes to have something like this, given that you know the puzzle and what you want the robots to do.
No self-awareness was there for sure, though, and I feel whatever they did won't really traspose to anything else easily, so I don't think the result is particularly impressive.
OTOH, the algorithm itself could be new and more performant of whatever Prolog could do, and that would be an achievement. But not really with respect to AI.
