> If we ignore compute requirements, how long would a good algorithm take to learn something like balancing a quadrupedal robot, based on real-time feedback to its outputs? It is a simpler problem, but drone flight software based on learning, can re-learn how to fly a quadcopter after something like losing a propeller and a sudden shift in weight distribution, in a few seconds.

Firstly, the drone software is not re-learning anything; when a baby walks unaided for the first time and you place a rattle in their hands, do you really refer to that as "re-learning walking"?

Secondly, forget about computational requirements and just look at how many iterations have to be run to reach success; mammals (and humans specifically) can "learn" to recognise an entire class of something just from one or two images of that something. Learning something else doesn't "forget" the old thing. A NN trained to recognise faces in images can't, to me knowledge, be trained to also balance a bi-pedal robot, or ride a bicycle.

IOW, what I see currently is that NNs are very one-dimensional - the net that becomes optimised for one task becomes unoptimised for that task if you subsequently add new tasks.