Interesting though that they seem to be using the demonstration only for initial states and not for action choices. It's like using an example of solving a maz

Interesting though that they seem to be using the demonstration only for initial states and not for action choices. It's like using an example of solving a maze just to get a bunch of places to start exploring from, but not to actually try and copy someone's "turn right at every corner" strategy. The use of domain knowledge is actually pretty limited in that sense..