[MagnusWootton]

Making encoders is a bitch, and dc motors are cheaper without them.

So it would be cool if u could get a robot to work without them.

Its a seemingly impossible thing,  cause a legged robot cant even have a centre point for its motors to even just stand still!    If you put mechanical end stops, then you can at least know the end point positions,   that could help.

If you dont have positional information, the trick is maybe u could correllate it from accellerometre information over time,   so you exchange the positional information for so many accellerometre samples and maybe its equivilent information!

So u just need to correllate it,  with machine learning.

Its actually a little of a long story to explain it,    but I think its possible!   U just need to get in the right semantical/generation space to make it happen.

So if you take it from the position of searching inside of a physics engine (as the generation space), you predict the next accellerometre accelleration, from a chain of accellerometre data, and u step it each new frame,  and the chain of accellerometre data is replacing the lack of angular position information, but it should be equivilent.

Then that should build the virtual copy of the physics, which lets u then brute force the robots motor commands to get the robot in action,  WITHOUT encoders!

[HS]

Wouldn't errors build up over time?

[MagnusWootton]

It essentially still has the angles of the motors,   it has them in the form of another variable that correllates to it,  being the accellerometre values over time predicting the next accelleration being in place of the motor hinge angles.   

but theres 2 catches.

u can only have so much size of input space on a neural net, before the function takes too many records to be saturated enough to predict the output successfully.

if u used the hinge angles directly,   its 16 hinge angles on a 16 hinge robot. (say a quad spider with 3d ball joint hips and a knee and 4 legs) -  so u get it in about 2^16 records. (every combination of high and low and then linear interpolation takes over for the rest.)   but that takes 2^16 frames to train enough of a set for the neural network to function properly.  thats 64k frames it has to learn.   if it were learning at 10 frames per second -  it would take about 1 full minute to get to 1% total perm saturation.

so it has to babble about for a minute to fill its database enough for it to start predicting the physics.

but if it were more than that,   say 32 hinge angles  -  it would be 2^32 frames to train enough for a functioning set,  thats 4 billion frames it has to learn,  if learning 10 frames a second it would take 49 days!!!   way too long.

so u can only have so many records before the records take too long for the robot to acquire them from its sensor data stream.   so im just saying instead of recording hinge angles,  why not just record an accellerometre over time till you get to 16 input values and just correllate it, instead of using encoders.

and it may work the same,   the accelleration is still predictable,  just didnt put angles inside the input of the network!

so the other catch being you dont actually have encoders, but ur correllating their "existance" without actually having them as an input.