Help requested: Improving the Calibration Process

how do the calibrations compare with a manual measurement?

1 Like

Thank you for your thorough report! I especially like the plots, they give a very good visual on where the calibration stands right now, both impressive and with room for improvement. Can you tell me what size the largest grid had?

Thanks, Arjen

2 Likes

@bar - thank you- I was struggling with a visual format that would both show the individual plots and the spread of the outlier(s). Given that each grid is 5mmx5mm and that except for the top two plot six each set fits within a 2mm circle, I’d agree it’s pretty impressive.

@dlang - taking the average of the first five calibrations’ final results’ calculated distances and comparing it to my manual measurements, it’s interesting that some are higher and some lower; but all fairly close.

bl<->tl tl<->tr tr<->br br<->bl bl<->tr tl<->br
Measured 2359 3578 2363 3577 4282 4291
Calculated from calibrations 1-5 2362.880194 3575.004511 2368.540482 3574.53999 4289.227454 4284.119019

@arjenschoneveld - All were done as default 1000mmx2000mm 9x9 calibrations on the hourglass frame which is roughly 2400mmx3600mm.

1 Like

entering those numbers into the manual calibration I get
Screenshot_20240429_094945

2 Likes

Which numbers?

entering the manual measurements gave the results above

1 Like

@dlang -Thanks for the clarification.

I don’t know if calibration is already skipping outliers, but would it be posible, as an extra step at the end:
for each measured gridpoint, calculate the deviation from the anchorpoints the calibration came up with, and skip the outliers (say every deviation thats higher than 2.5mm), and run the math again with the remaining numbers.
I thought of this when I drew the 4 cornerpoints of the largest grid of the 6th calibration of @CPL in CAD , the deviation of 3 of them was less than 2.5mm, the top left gridpoint had a deviation of 4mm, so I think something might have gone wrong there during the calibration.
Maybe only do this last step when the initial fitness number is higher than 0.6?

Arjen

2 Likes

That’s essentially what I am trying to accomplish- I want to run a series of calibrations to generate a larger dataset, and see If that will yield a usable -“better”- result.

1 Like

another thought: A quick way of checking if each encoder or arm measures the same distance would be to hook 2 oposite belts up to 2anchors, then measure, then unhook them , and hook up the other 2 belts to the same anchorpoints and measure. If the measurements differ, you know that something is wrong. for instance: the tension at which the measurements take place differs for the same amperage, which would mean more friction in one arm.

Arjen

1 Like

I hope it will, and the bigger the better I think. That only takes a lot of time, I was trying to come up with a way to shortening the process by getting rid of outliers in one calibration proces.

1 Like

I don’t know if calibration is already skipping outliers, but would it be posible, as an extra step at the end:

No, currently it does no sanity checking of any of the datapoints.

for each datapoint, it gets the lengths of the belts, projects them from the 4
corners of the guess to a crossing point, figured out which one is the furthest
wrong, and changes the guess for that line to make it less wrong,

David Lang

1 Like

arjenschoneveld wrote:

another thought: A quick way of checking if each encoder or arm measures the
same distance would be to hook 2 oposite belts up to 2anchors, then measure,
then unhook them , and hook up the other 2 belts to the same anchorpoints and
measure. If the measurements differ, you know that something is wrong. for
instance: the tension at which the measurements take place differs for the
same amperage, which would mean more friction in one arm.

unless belts are stretching, the encoders are all going to match over the long
run.

they have toothed gears that mesh with the belt and have absolute magnetic
encoders to measure the angle of the toothed gear. It can’t ‘miss a step’ (it
could, in theory miss an entire rotation, but that’s something like 40+mm) and
the belt isn’t going to slip on the gear due to the teeth.

now, there is a little room for error, do repeated retract/extend cycles and you
will see that the retraction doesn’t always move to exactly zero

David Lang

yes, but if belt tension is measured by looking at the motor current, I think there is room for error there, as we see people needing to up the motor current to retract all belts. It is probably a small error, but we are talking about small deviations here.

I can’t check that because I haven’t got my machine yet, but that seems strange.

Arjen

1 Like

arjenschoneveld wrote:

yes, but if belt tension is measured by looking at the motor current, I think
there is room for error there, as we see people needing to up the motor
current to retract all belts. It is probably a small error, but we are talking
about small deviations here.

belt tension is only measured to decide if the belts are tight enough. It does
not use tension to calculate belt stretch.

I can’t check that because I haven’t got my machine yet, but that seems strange.

When you do a retract_all it pulls until the belts top moving (at which point
the current applied to the motors goes up in an attempt to keep them moving at
the same speed, until it hits the current limit and decides it’s fully
retracted). At that point it prints out what it thought the length of the belt
was and sets it to zero.

We should probably have several people (including those who have had to increase
their current limits) do several retract/extend cycles and report the stopped
lengths.

David Lang

1 Like

so maybe I’m overthinking it, but if you have 2arms, one retracts fully at 1300 mA but the other one only retracts fully at 2000 mA due to more friction in the hardware, if the “tight enough” trigger is set at 2000 mA for both arms, the tension on the first arm belt will be higher than the tension on the second arm belt. The higher tension leads to a slightly longer belt because of belt strech. This effect should be very small because of the steel wires in the belts, but it could be a factor.

O and I’m really not trying to say the design of the machine is bad, I think it is awesome. I’m only trying to find points where a source for mismeasuring could be, so it could be checked, and if needed corrected in software.

Arjen

1 Like

So you retract , and define zero, then extend, then retract again to where the machine thinks is zero, and there is a gap?
Or is it not on the zero as it ends retract, just triggers the motor current trigger?

arjenschoneveld wrote:

so maybe I’m overthinking it, but if you have 2arms, one retracts fully at
1300 mA but the other one only retracts fully at 2000 mA due to more friction
in the hardware, if the “tight enough” trigger is set at 2000 mA for both
arms, the tension on the first arm belt will be higher than the tension on the
second arm belt. The higher tension leads to a slightly longer belt because of
belt strech. This effect should be very small because of the steel wires in
the belts, but it could be a factor.

you are correct, but the system does not (currently) try to take this into
account

David Lang

1 Like

arjenschoneveld wrote:

So you retract , and define zero, then extend, then retract again to where the
machine thinks is zero, and there is a gap?
Or is it not on the zero as it ends retract, just triggers the motor current trigger?

It is not on the zero as it ends retract (which is ended due to the motor
current trigger)

It then sets that position to be the new zeros for the belt lengths.

David Lang

can you see how much the difference is?