Here's some thinking out loud about walking and the steps needed to program this or "a" numerical structural system to make this work. A strategy.
So as I walk about, I started thinking about what I'm doing. It appears, to me, that walking is sort of a natural preprogrammed act where the muscles mostly know what to do. Brain tells body go here, go there, it goes here, goes there. So I started thinking about how we could do the same.
I came up the idea that with small amounts of data passed to the muscles we could get good movement.
My assumptions are we are using micro-controllers to control each muscle, with some extra processing power for each muscle.
The first step is the brain makes a "map" of the terrain in front of it, what speed it is going at and then where it needs to step for that speed or terrain. I watched a Jim Keller interview with Lex Fridman and Keller said that this sort of distance measurement was trivial. In other words using two eyes you can tell where something is like the floor and how far away. He should know, he's a legend at AMD and Tesla, designing chips for...everything. So let's take his word for this, (yes there are likely to be complications but let's ignore for now). We will assume finding a place to step for the waifu is not big deal. A simple table of sorts could be built up of how far a step is depending on speed it wants to go and using that it maps where a good place to step is.
So now we have that, then, we need a vector to this place. Now a leg, thigh, foot, movement is going to be constrained to how they can bend so if the brain sends a start moving vector(a angle for the foot to move AND a velocity), a end point to go to, then this set of vectors from the brain could be interpreted by ALL the muscles in the leg. Each one knows that to move here or there, it must act in a certain way. So with this one vector based on, move this way x, this way Y and this way Z, plus a speed, plus the end point, each muscle can work on its on to add up to this end point. So you have,
(Each 2 bytes(16 bits) giving us 65,536 different values or postions)
A starting direction for foot movement, x,y and z. If you need to step over something it will have a high value for "Y". It will pick the foot up high.(6 bytes)
A velocity point for x,y and z, how fast to move(two bytes x, two bytes Y and two bytes z) (6 bytes)
End point (two bytes x, two bytes Y and two bytes z) (6 bytes)
and a foot position x and y for how the foot needs to land on the ground.(4 bytes)
(22 bytes total)
This will create a vector that tells the foot, leg, thigh, hip how far it is to lift the foot, how fast and what direction x,y,z.
This data is really just where the foot moves. What direction and how fast to move in that direction, where it is to move the foot to and the position of the foot when it lands. All the muscles have to do a certain task to do this and they "know" what to do, they do their part by only feeding them foot movement data.
Another thing I think that would be needed to give the waifu a real strong grasp on what to do, is to feed the muscles a body position based on the body mass point at the shoulders and the hip AND a velocity vector x,y and z of what the hips and shoulders are doing in respect to movement. I think these two measurements could really add a lot to it's position awareness with a small amount of data being passed. In this case.
shoulder x,y,z position(6 bytes)
shoulder x,y,z velocity(6 bytes)
hip x,y,z position(6 bytes)
hip x,y,z velocity(6 bytes)
(two feet, 40 bytes x 2, so 80 bytes total)for feet and body in all.
At 50KBs with 80 bytes, let's add in some error correction and maybe some other stuff so make it 500 bytes just to throw a number at it and you could do this a 100 times a second or every hundredth of a second. That's fairly good at a really slow network speed, though I would like to see the time interval smaller and kick the transfer speed up a little.
It worth noting that if each processor can figure out it's moves on this small set of data it doesn't need continuous data streams. Only one set for each step. The walking models, that say they are successful, say they add refinements for the last part of the step so you could send a start and then small refinements as it continues through the step with way, way, less data than 500 bytes every 1/100 of a second.
There's also two paths that could be followed. The main brain could figure out the proper step pattern for a specific body position and just send the 22 bytes foot movement step, velocity vector and end point to the muscles to walk. With each muscles processor figuring the actuation needed to preform that specific movement. Or it may be that some combination that changes depending on circumstance is used. Where sometimes the muscles get body data, (the shoulder, hip), positions and sometimes they just get foot movement data.
I was pondering this and just decided to jot down what I was thinking.