Calin said:
Have they gotten any further than that? Since we have vaccum robots finding their way around the house, one might think robotic bipeds should at least be able to achieve the same.

The problem of controlling a biped robot is harder than what a vaccum robot does, so if they wanted, they surely could build expensive biped vacuum cleaner robots. The question only is if there's demand. For most tasks it's just more efficient to build specialized robots just for that. An universal biped robot could do many tasks, but at much higher costs.

Which brings us to the big question: Do we even need biped robots? I'm not sure.

But i think we need those control systems for games. So we can get rid of expensive motion capture which still produces only static content in the end, and to increase dynamics in games beyond the very simple and restricted models we use now.

Lots of interesting problems in that space, but no revenue. It cost about US$120 million and 30 years for Boston Dynamics to get to Big Dog. Then Google bought them, and put in even more money. Then Softbank bought them, and put in more money. Now Hyundai owns them. They sell a very expensive robot dog.

I used to work on legged running. Back in the 1990s. Figured out how to do the equivalent of ABS for legs. Turns out that anti-slip control dominates the problem once you get off flat surfaces. Balance is secondary.

The Google acquisition was not a good thing. Google bought up several robotics companies and accomplished little or nothing with them. One of the good ones was Schaft. Nobody bought them, and now they're gone.

Calin said:
To my mind once youve got walking in straight line and balancing figured out (which they did) it`s a matter of getting the robot to follow way points to get him to move around and avoid obstacles.

I did this with the one legged hopper, but not yet with the walking biped (hopper follows the blue target):

Recently i have started to maintain and continue with this old project a bit. Developer of Newton engine also works on self balancing ragdolls, so anyone interested might find an easy to use solution there (soon - it's wip; primary idea seems to reuse physics constraint solver for IK).

Maybe interesting: Such one legged hopper is no good example to work on balancing. It can maintain unstable balance only temporally and has no control over it. It's probably useful to derive a model of running bipeds from it, while a walking biped is a completely different problem allowing precise control over stable and unstable balance.

After primary locomotion works (walk, run, climb, swim…), the harder problem surely is interaction between multiple characters. A simple inverted pendulum model can no longer represent this complexity, and likely we need some nervous system reacting to the inputs from physics contacts.
Then we have navigation through constrained environment, e.g. crawling through a narrow cave between rocks. To avoid collisions or at least injury, we would need to solve complex many body 3D pathfinding problems. Not practical. I hope nervous system with local response and adjustments is good enough, and this seems to be what real humans do as well. ML might be the best way to implement this, but personally i won't go there.

Nagle said:
I used to work on legged running. Back in the 1990s. Figured out how to do the equivalent of ABS for legs. Turns out that anti-slip control dominates the problem once you get off flat surfaces. Balance is secondary.

I remember this video - very nice! I have ignored friction til yet and it's not handled in my controllers. Not sure it should dominate over balance. My plan is to keep ignoring it, and controllers just assume friction is always high enough to prevent slip. Then, e.g. if walking on ice, a reflex system should model a quick reaction to slippage preventing tipping over. Will see how that works out…

Nagle said:
Lots of interesting problems in that space, but no revenue.

Recently i saw Natural Motion no longer shows ragdoll middleware on their webpage. So it seems hard to make this a success for games too.
I guess it was not really easy to use, but i'm still a bit disappointed from innovation lacking games industry not picking this up. Everybody loved their videos, and i remember it generated a lot of fun in GTA 4.

That would probably be the last stage in developing a universal biped robot. Getting the biped to vaccume the floor is the first challenge that needs a solution I would say.

Well, if you build a biped cleaner, it will operate in human territory, so you need to model some interaction with humans anyway, if only to prevent doing harm to nearby humans.

> To achieve that you need some type of universal problem solving algorithm.

Yeah. Can you give me a link to a paper describing such algorithm? I'd really like to have that :D
Probably the closest we actually have is ML. But i think we are not there yet, accepting such unpredictable and fuzzy solutions to be fundamental parts of our games.
We want as much determinism as we can get. So i'll try to subdivide big problems into smaller problems, solving them with simple and predictable controllers as far as possible.

Getting a biped to do stuff in a real environment would of course pose one additional challenge: transforming the the real world into a mesh/FPS level (First Person Shooter level).

Exactly! This is why we should solve this in games first. We have accurate sensors for free, so a big part of the problem does not come up for us. This is not only about world geometry, but also about sensing velocity for example, which is hard in the real world.