We actually do have many things along that line, motion capture for 3-D movies instance comes to mind almost immediately. The problem if I think about it is less of a situation in observing another actor, computers are relativity good at doing that already with the amount of image recognition software we have, rather it's a problem of understanding if an action yielded a good outcome as a net which is something that computers cannot do as it's not a single node network problem. For example, we've already programmed a computer to understand human language (Watson, arguably), but even Watson didn't understand the concept that saying "f***" is bad. (Look that up, it's a funny side story.)
But the point is, learning algorithms are not true learning in a sense as a computer currently has no sense of "a good outcome", hence at this stage observation learning is very much limited in a sense to "monkey see, monkey do".
Perhaps the closest thing I have ever read about with this was firefighting search and rescue bots that were on a network and would broadcast to each other when one of them had been destroyed as the bots would know the area was something that they had to avoid.
Otherwise, I think this is the problem with observational learning. A person can observe that punching someone usually will get you hit back, a computer will observe and parrot the action, good or bad.