I worked with Atlas during 2013 at WPI. Siyuan Feng just defended his thesis on Online Hierarchical Optimization for
Humanoid Control for controlling Atlas during the DRC. The controller used is not as well refined as BD's controller, but I believe it stands in the same principles, and relies on ZMP.
Short answer: No, ZMP is still the name of the game +-. [0]
Slightly longer answer: No, they use derivative approaches that are basically ZMP but generally more refined/modernized, which is not to say ZMP/related are good or the 'future'. [1]
Source:
[0] I can try and find some non-paywall PDF's that go over the field, but basically everyone uses ZMP except for a few research groups in Korea or UMich and groups using their design/model.
Disclaimer: I don't work with these types of bipeds. I work on powered exoskeletons for paralyzed type people. I would google that phrase to get an idea, but what you'll find is not close to what i'm working on (Read: Wink wink). That being said, the fields overlap very heavily, probably too heavily given the necessary differences, so I do have a fairly rigorous (rusty, at this point) understanding of the topic.
You are right in a way; big dog et al (depending on specific bot/type) share alot of software features with biped robots, but not the type that deal with gait control. The difference is found primarily in gait-control/kinematics, as determined by the necessarily complex nature behind bipedal mechanics.
The future of robotics sort of already exists amongst us, at the very least in the public research domain. The fundamental issue with biped robots right now, as i see it, is:
1) Biped robots are almost universally 'human' inspired
2) Humans are theoretically fully optimized bipeds, copying them is logical and probably an unbeatable 'design'
3) Designers try to copy the human form, but either focus too heavily on ZMP influence or too much on a 'Terminator' level robustness (Heavy, rigid, and impossible).
4) The robot is produced, and because ZMP is very predictable/safe, it walks; but it is not 'human' form or very good at achieving it's design goals because its goals were modeled after the ideal biped, and this is a very bad imitation.
5) Designers say: Hey, spec sheet checks out, FEA says this is strong, problem must be bad software [ZMP]. Software guys say: This software [ZMP] can be improved. The designers are right, the software guys are right, but ZMP is wrong, and because of that, design is bad, etc...
So to really get a true, true biped robot, you really, really need A) Control theory that isn't backwards [literally] and B) A really, really 'human-like' (musculoskeletal) design.
A and B are both really difficult because both are massively challenging, and they have to work together perfectly. So you need a solution that is really creative, really beautiful and elegant and lightweight, and probably most difficult, you need to have people that know how to turn that kind of framework from theory to iterative simulation to real world etc etc.
ZMP is the only real game rn because it was first and etc [sunken cost fallacy and myopia, i think]. All other approaches are neglected fully, regretably (not many other approaches either).
Jumping back to how I see it, I think bipeds will only improve superficially till ZMP is killed or hybridized effectively, which doesn't mean A and B has to occur. Their are just too many fundamental limitations to the theory for bipeds using it to become ubiquitous, or even used beyond research/darpa-fairs.
I also think that my research/focus is more complex than robotic bipeds, and it'd be really worthless waste of my time if it was at the current or slightly above 'state-of-the-art', but this is absolutely the best way I can spend my life right now.
This whole industry [true 'wearables' and robotica] is like a giant powderkeg, and it's going to blow up way bigger and sooner than any thought possible [imo, so take with a grain of salt, but facts afford a potentially representative state of things to be].
I've also just realized i've forgotten to explain zmp/gait and many other things, but i feel that the write up stands on its own.
Genuinely curious: could you chance a guess as to why they chose bipedal over quadrupedal?
I don't know much about the field, but mechanically quadrupedals simply make more sense - a lower center of gravity combined with more opportunities (in the form of more legs) to correct blunders.
I don't think traditional ZMP would enable the stuff this robot does in the snow. More likely some sort of evolved controller, like in this paper: https://vimeo.com/79098420
This is a textbook, ZMP derivative type controller.
ZMP doesn't necessarily require absolute maps of the environment, but generally is more of a predict and optimize for safety type affair, and Atlas i'd imagine does that both very quickly, and a bit smarter leg reaction movements.
The linked video is more akin to elementary central pattern generation, which is a very understudied field, but their is more advanced research on the topic from pre 2010 if you are curious. Interest seems to have dropped off unfortunately in the really cool stuff since then, but you can find it with moderate ease.
It doesn't seem to have a very natural gait. Very uncanny valley.
https://en.m.wikipedia.org/wiki/Zero_moment_point