Comment by Animats
> Boston dynamics is a leader in getting robots to do useful niche work in well bounded environments, but that's yesterday's news.
BD did most of their locomotion using classical dynamics and control theory until a few years ago. So did Honda, with Asimo. I did some of that in 1994.[1]
Early thinking revolved around landing on the "zero moment point". There's a landing point which, if hit, maintains speed and balance. To speed up, you aim for slightly beyond that point; to slow down, aim for a nearer point. That was Asimo. You could push that concept to the level of BD's "Big Dog", and later, their smaller dogs. Even pre-calculated flips were possible. But that approach gets you rather clunky motion.
The next step was to use some machine learning to tweak the control system parameters. That works, but you don't get overall coordination of all the joints. That only started to appear as machine learning systems became powerful enough to take on the whole problem at once.
Hard problem. Took over three decades to get decent humanoid control. Now everybody is doing it. You can be too early.