The tough question is how exactly to do this... I think you could do 1 and 2 by hand at the start. Skill 3 is where the next advancement should be. Making a neural net of neural nets, perhaps...

There you might (in step 3) run into real complex problems. Throwing while running might to too different to assemble from throwing and running that you would get absurd behavior like running, dropping like a brick, stand up and throw a perfect ball (kinda like children do).

Some interaction within the learning process by humans might not only speed things up but also prevent some comically disastrous solutions. But this won't be cheap and there is the "out of the box" solutions the system may come up with that are never reached because the humans interrupt it when it wants do to a flip {still the baseball example}.

Learning amongst humans isn't that well understood so hopefully these two problems wholesomely help to solve eachother.

And if it all works like we want it to, in a time far far away, some ***bag would go "Can we MKUltra this system?".