Where do you see it learning to walk? You can't equate the inverted pendulum problem with upright walking. It is neat if it can choose it's own reference level ("goal") for a classical control problem -- but that's a long way from "learning to walk".

I'm far from being blown away. Researchers in AI have demonstrated AI systems with a "blank mind" that can learn by interacting with its environment given some reinforcement learning. They often operate in simple toy examples, but once confronted with the real, complex, messy reality, they fail. I found mostly two reasons: Either their mechanism is too simple and can't solve other than simple problems or it's (computationally) too complex and doesn't scale.

So I'm not saying this approach is bad, but we have yet to see whether it does more than simple examples.