DL mainly focuses on learning from data. So that will be very challenging for the current state of technology. But for RL, one objective is to learn effectively as human. So the first question is: does human perform well under this limit set of data? If not, then AI will have a tough time also. But being said, it is still under heavy research of what does it mean in “learning effectively”. There is some progress to teach robots how to perform tasks under two hour of training. Google has done a lot of research in self-training also. i.e. how to learn with as little data at the scale of human-level or just learn things without human involvement. But we are still not there yet and the tasks that can perform is like picking arbitrary objects with a robotic arm.
I am not sure what is the definition of “non-functional” physical system. So I have no comments on that.