It took Boston Dynamics about 15 years and $125 million to get to Atlas. There's still no market. This business requires a sugar daddy. BD had DARPA, then Google, and now Softbank. This is a very expensive area in which to work.
Robot manipulation in unstructured situations works very badly. Watch these videos:
DARPA-funded robot manipulation work from 2012.[1]
DARPA-funded robot manipulation work from 1973.[2]
See the improvement? No?
It's not clear that just throwing deep learning at the problem will work. It's been tried. Even general bin-picking still doesn't work.[3] The author of that article, who is from Fanuc, says maybe by 2020. Lots of special cases work already, but if all your parts are the same simple shape, a bowl feeder is simpler. Picking from a bin full of partly entangled parts is beyond what robots can do today.
Someone got bin picking to work for an irregular sheet metal part by going in with a magnet, picking up something, anything, then weighing it. If it's underweight, they missed, and they try again. If it's overweight, they got more than one part, so they drop it and try a different location in the bin. If they got one part, they now have it out where they can look at it and orient it. Tricks like that make production lines work.
Pure geometry isn't enough, and pure machine learning isn't enough. But together, they might work. Good to see this being funded.
Here's the 2017 Amazon picking challenge.[4] Still not there, but better than last year.
[1] https://www.youtube.com/watch?v=jeABMoYJGEU [2] https://archive.org/details/sailfilm_pump [3] https://www.robotics.org/content-detail.cfm/Industrial-Robot... [4] https://www.youtube.com/watch?v=1QqQLq5hsN4