Ai Dreams Forum
Artificial Intelligence => General AI Discussion => Topic started by: ranch vermin on September 25, 2018, 03:15:45 pm
-
This is hard to explain, i dont expect anyone to get it exactly, but its a cool new idea.
I wonder if you can make the robot learn its spacial posture, instead of using gizmos, accellerometres or potentiometres or step motors - or any kind of feeling sensation, or motor force feedback.
Every time you move, you add both a motor difference key (which doesnt require posture, but requires a little over time to match up to the movements... and it also requires information from the outside, to properly match to itself.) and a posture guess.
In the database, where all the motor differences match up, you can guess the postures are similar here, so you get to reduct possibilities, and the posture guess result, is a direct error to its database, so collect both of these over time, and u might be able to make a robot more adaptive than using potentiometres or step motors, which are not trustworthy once they are damaged.
-
OK, seems like a good idea to replace potentiometers and such to make the robot more robust. How are you gonna measure motor difference?
To answer your question about the 3D print, the material is nylon.
-
That sounds like a type of ‘dead reckoning’ to judge limb positions. I don’t personally think it would be accurate enough and you will still need a reference/ calibration point.
I don’t want to encroach on your project but we humans use our visual system to fine tune/ calibrate our body posture.
What about a HD camera literally on a stick, looking down on your quadruped. You could use reference/ angle/ markers on the limbs to calculate their positions and also get a good view of the terrain around the bot... cheap & easy.
:)
-
yes, cameras can do it. i imagine you could get away without a stepper motor on your 3d printer if u hook a cheap webcam on the nozzle!
I thought about it more, and ive got an easier way to explain it. U may already know what im going to say, because i dont know what "dead reckoning" is, but is it this?
Ive got a left charge and a right charge known every frame, so its a matter of taking the memory fragments and then piecing the lefts and rights over time and it could craft out the continuous posture out of the binary directions, constrained by how they have to meet at their joining locations.
-
All good ideas but…
Any type of position sensing without an index point to set/ calibrate 0 positions, is prone to drift. A Geared DC or servo just isn’t accurate enough, besides gear slap you have inertia/ overrun to contend with. Even a well balance/ loaded stepper can be moved if enough force/ torque is applied.
I personally think these small errors would compound in a very short time period.
:)
-
Its crazy :D, thinking i can just run a sentient humanoid just off the fly like this, but I think it might be worth a look still.
All I said was less or more, its actually true it was less or more, as long as i dont constrain anymore than that, no more constraint is given, until more information is provided, then
its when the memories link to each other, where they are considered to have to be the same posture, or postures after each other, where then i can start knocking off possibilities.
[EDIT] oh i read what u said, if the arm is moved FOR it.... it wont be able to deal with it. [/edit]
-
I was wondering, if load is detectable on a motor, because that may be all it needs - to detect this "opponent force" maybe u could even form a gravity vector from it.
Might end up just as troubling as adding the potentiometre, just a different technique.
-
Even if you're trying to do something as simple as getting a wheeled robot to drive a specific distance, you want sensors ... unless your desired margin of accuracy is really wide. Trying to manage a complex multi-jointed system without any sensory feedback, and non-stepper motors, sounds like a nightmare.
-
https://www.youtube.com/watch?v=v4BbSzJ-hz4
:)