- teleoperation is generally orthogonal to creating a robust bipedal gait, and it seems like capture point approaches by BD & whatever Shaft is doing, if one can believe their marketing videos, as well as underactuated systems like DURUS and ATRIAS (and its bird-like successor, can't find the name, C-something) are starting to look quite robust. Some work still to be done, I think but looking quite promising. And this is all just rather classical approaches; nothing machine-learning-y at all. You'd still teleoperate them to navigate and make them do a manipulative task, but not just because it can't walk otherwise; but because those are different domains and problems (like manipulation, autonomous navigation etc). Hell, ATRIAS is completely blind anyhow, all you can do is define foot clearance height and direction of movement.
- deep reinforcement learning based visuomotor control techniques demonstrated last year on the PR2 platform and later on I think also demonstrated on a Baxter robot seem pretty dextrous at such manipulative tasks like folding shirts, and does it realtime, after some trial and error self-training. Granted, didn't see them do exactly that task, but a bunch of other manipulative exercises were demonstrated, realtime and robust to environment destractions, perturbations etc.
Admittedly this later part is still limited in that the problem needs to be fully observable at all times by the current robot -- it can't deal w say occlusions for it has no memory of what it previously saw -- and its not actually trained to be robust to disturbances and distractions -- what it demonstrated on that front is essentially just what it gets for free from the robustness of deep approaches -- so it could be improved.