It was kind of disappointing to see the transition from fancy, glamorous shots of it standing still and looking around/at its hands to plodding slowly and cautiously (I was kind of reminded of walking in ski boots?) across a flat, open floor.
I understand that bipedal walking is hard, but I don't really understand why. Is it a materials thing? Do you need the right kind of stretchy tendons and ligaments to produce more fluid motion? Is it a sensor thing, about having the right kind of signal to inform control for dynamic balancing? Is it about having a lot of different degrees of freedom in your toes/feet/ankles to achieve balance, which aren't yet feasible in robots?
Googling for "bipedal walking is unstable" returns a plethora of articles whose abstracts suggest that yes, bipedal walking is physically unstable. To quote one [0] of them:
Robust walking on two legs has proven to be one of the most difficult challenges of humanoid robotics. Bipedal walkers are inherently unstable systems that are difficult to control due to the complexity of their full-body dynamics. Aside from the challenge of generating a walking motion itself, closed-loop algorithms are required to maintain the balance of the robot using foot placements and other disturbance-rejection strategies.
... I do think "it's unstable" is kind of a non-answer though. Lots of control problems are unstable, right?
For example here's a successful project from 16 years go where Andrew Ng, Pieter Abbeel etc used apprenticeship learning to teach a controller to fly RC helicopters through complex stunts, and they discuss how these flight patterns have to deal with instability, and the controller is eventually doing things very fluidly which humans would struggle to do.
So what about bipedal walking is harder than other learned control problems that RL can master (and could handle even on previous generations of tech)?
The idea of using humanoid robots in the automotive industry has always been a bit confusing to me. Automobiles are all the same within a model and are built on an assembly line, so they are well served by robots with fixtures specialized to each production step.
Wouldn’t a humanoid robot be better suited to factories with batch processes or engineered to order products?
Would be curious the hear the perspective of anyone in the automotive industry - do you see humanoid robots being useful in your industry?
Well, nobody can answer that question because humanoids don't exist. That is, the only way to get that answer is to actually try it out.
Broadly speaking, humanoid robots like this should be the "worse is better" kind of tech. They're cheap ($20-30k at scale or like $1/hr), imprecise (cm vs mm), weak (manipulate several kg) and flexible. The question is where in the factory can you take advantage of this kind of labour or can you redesign your process to take advantage of it.
I understand that bipedal walking is hard, but I don't really understand why. Is it a materials thing? Do you need the right kind of stretchy tendons and ligaments to produce more fluid motion? Is it a sensor thing, about having the right kind of signal to inform control for dynamic balancing? Is it about having a lot of different degrees of freedom in your toes/feet/ankles to achieve balance, which aren't yet feasible in robots?