TCP orientation problems

[wes_mcgee]

Ok,
   So I am running out of ideas here. I have a KR100HA, which is considered a pretty accurate, and repeatable robot. I have a Dynalog Autocal system, which supposedly guarantees to measure teh TCP of a spindle like tool to within .5 mm(though I am not sure of the orientation accuracy numbers). I have also used the loaddetect software to get an accurate COG which kuka absolute accuracy uses to compensate for load deflection
   I wasn't getting the results I wanted, so I had the robot laser calibrated by Kuka with a FARO....they certified under .5 mm across the whole envelope. To make the long story short, after FARO calibration, my autocal measured TCP orientation was significantly different. This could be OK, I am thinking, and everyone agrees the proof is in the test

Here is my actual test setup. From a fixed position on my external axis, I am using an 80mm tool to machine a plane 1.2mx2.4 m, parallel to the worldXY. I had done this before the AA calibration, and had less than good results, so I was eager to try after the AA calibration and subsequent recalibrated TCP by Autocal.

Here are some assumption, feel free to discuss these. I am looking mostly at local flatness, IE if the tool steps over 60 mm(at the same Z height and orientation, how "flat" is the surface. Over larger differences I realize the load deflection could under or over compensate, causing the surface to slope up or down slightly. But locally this test should be all about proper tool orientation with respect to the worldZ

I wouldn't be writing this if the test was good. It was worse than before the laser calibration. To test, I manually(visually) squared the spindole to a surface I knew to be close(no, I don't have numbers for this) to the world XY. At least it was close before the AA calibration. It was visibly off, so I manually adjust the orientation values from the tool until what looked square was indicating square in the rob position readout. This is just a test, I know we are rounding here, but it was off by 1.5 degrees in A and C, so that is pretty visible with a large diameter tool. I then reran the planing job, and yes, it was much more flat locally. Not perfect everywhere, and that is because in order to run the program in my workcell I have to vary the spindle rotation (around the tool, Z) gradually over the whole program. That would work fine with a perfect tool orientation, but we don't have that here. The fact that I got a "flatter" surface by eyeballing the tool orientation that with my $$$$ calibration device has me more than a little !

My next test is to get an actual spindle tramming setup(circular parallel, dial indicator in the spindle) used on a milling machine, and try to get better numbers. I expect the process will be to use a properly mounted, TCP calculated dial indicator on the arm to square the circular parallel to the worldXY, and then use the dial mounted outboard in the rotating spindle against this parallel to iterate the values until I get the TCP aligned with the worldZ(my working direction is Z).

So the question is, am I missing something here. I have contacted both vendors(no reply yet). Does anyone have any other good tricks for getting accurate TCP orientations? I used to love the 4 point plus Z method on ABB robots, it seemed very accurate. Before I pay anyone else I think I would just rent a laser tracker on my own, it should be pretty easy to find the axis of a cylindrical tool with one(we weren't able to do this at the time because I couldn't load a tool with the dresspack removed, which they required for the absolute accuracy calibration).

thanks in advance, sorry if this isn't clearly articulated.

[SkyeFire]

Just out of curiousity, what kind of accuracy are you trying for, and what are you getting?

It sounds like you had more than one issue going on before, but am I correct in my understanding that your local flatness is better now?  Within tolerance, hopefully?  So right now your biggest issue is that you need a really good TCP, including orientation?

The KUKAs have a similar method to the ABB 4-pt tool creation, but since it sounds like you're trying to achieve near machine-tool accuracy, I doubt that'll be good enough.  The best method I know of to get a really good 6D TCP value is to use a laser tracker to establish the 6D zero-tool (the robot mounting flange), and then measure the actual physical TCP in 6D using the zero-tool as the frame of reference.
You're still going to be vulnerable to things like tool flex, and robot flex over long reach.  This technique works best if you have the robot and tool reasonably close to the orientation that you'll be using during production.  Minimizing the amount of tool orientation change during your path also helps.

You may also be getting some flexion from your 7th axis.  I'm guessing that the robot is resting on top of a linear E1 axis?  I've found in the past that, as the robot reaches further out from its mounting point, even the "heavy duty" external axes will flex to some degree, letting the TCP "droop".  And while I hate to say it, if you have to move the robot on E1 for the same work piece, you're unfortunately likely to find that the machining you get from the different E1 positions are probably going to show some differences.  You might get lucky that way, but if not... hm.  Maybe (and this is a wild shot) you could get your E1 track or your workpiece tooling shimmed so that one is Really Really parallel to the other, and then use your E1 motion as part (or all) of your Y-axis machining path.  But that's probably for the future, so I'll shut up on that now.

So, my next question is, for this application, do you need accuracy or repeatability?  Because, even with a "rough" robot, you could probably achieve overall repeatability under .5mm (potentially quite a bit lower, given some of the test data I generated).  But to do it, you'd need to do something like use your laser tracker to "grid out" your machining program -- move the robot through a "grid" of points covering your machining area while capturing robot position vs tracker position, and then use the data to build a correction table.  A lot of fiddly work, but once it was dialed in, it should work great for a good long while.  I know the robots are capable of it mechanically.

Also, you could try something like this:  use the laser tracker to determine the bottom dead center of the robot.  I'm not directly familiar with the 100HA, but usually on KUKAbots it's the centerline of A1 rotation where said line intersects the plane of the bottom of the base casting (the machined surface with rests against the mounting plate and through which the mounting bolts pass).  There may be some machined locating pins as part of your robot mounting that you might be able to use to establish World X or Y, and find the centerline.  Or you could try tracking the robot in space and do a best fit, similar to the approach immediately above.  Once you've done that, put a workpiece that's already been finished and that you know is "golden" into the working position, measure it with the tracker, and try to establish a Base frame for the work piece from that data.  Then do the "grid sampling" as mentioned above (keep the robot close to the actual machining path while doing this for maximum accuracy), and see how well the robot holds parallel against the plane of the work piece.  Finding the overall slant, or the "lumps" in the grid data, might tell you something valuable about what you need to do.

Side note:  when your Absolute Accuracy calibration was done, did they calibrate for the whole robot motion envelope, or just the working envelope that you're using?  B/c I can say for certain that the smaller the volume, the better the calibration is -- "stock" Abs-Acc from KUKA calibrates the robot's entire physical motion volume, and averages the results overall. 

Another thing:  what does your machining path look like?  Articulated robots are, sadly, still very vulnerable to "backlash" issues, which even effect robot repeatability.  I've done this test -- run one of these robots to the same point, from the same starting point,, over and over, and I've seen repeatabilities as low as .1mm.  But come at the same point from another direction, and it's a crap shoot -- you could theoretically see a worst-case difference of something like 1.5mm.  I never saw it get that bad, but it was bad enough.  (I should mention here that, in straight-up mechanical accuracy tests between equivalent models of different brands that I either saw or had some participation in, the KUKAs generally worked as well or better than the competition -- this isn't a brand-specific thing, it's endemic in all articulated robots to some degree).  If your machining path doesn't take this into account and use certain techniques to minimize the effects of axis backlash ("lost motion," technically -- everyone insists that the gearboxes all of the major robot brands use are "zero backlash"), your machining accuracy could be taking a hit there, as well.

Hope this helps a bit.  Let us know what happens, will you?

[wes_mcgee]

Lots of good points there. Would love to try them , wish I could afford my own tracker! And maybe I will rent one sometime just to do it. It will take me a while to digest that. I do know they calibrated the entire envelope, so yes it is averaged, and yes, at .5 mm max error that is an issue. Currently I really think my current problem is an orientation one. While I don't think this is a backlash issue, I need to try "machining one-way" to try and eliminate that variable, because evidently axis one backlash would be the highest, though I guess 2 and 3 could be major as well. I am actually machining a large flat surface, yes from a fixed position.

As far as the question of "how" accurate, that's a good one. I don't really have a good answer. was hoping to be around .3mm global flatness over this 1.2m x2.4 m surface. The local issue, I need to test more. Basically the spindle is not running parallel to the world. A little trig says that a .1 degree tilt from the world Z will cause a .14mm step to appear with an 80mm tool at max stepover. Of course I could reduce that. It is just a rough goal; .14mm locally would be great. I am seeing more than that, I need to quantify it though. I am getting a tramming fixture to help. I have since talked to the vendors so they are helping as well.

I will let you know how things go.

[SkyeFire]

Yeah, those laser trackers are great toys, aren't they?  And several times more expensive than a full-up robot, too, not to mention the fact that you need a good operator to really get the most out of them.  One thing I learned the hard way, if you're renting one (or just billing one inside the company), have all your ducks in a row before the tracker gets there -- sometimes just grabbing a whole whackload of data in one big session, then taking a few days to process the data before bringing the tracker back for more focussed, interactive/experimental work is the way to go. 

And having a tracker operator (or someone else) who Really Understands the underlying geometries, rather than just how to use the tracker, is a major godsend.  There's a percentage of tracker operators who don't know much more than how to collect data with the system, because that's all they've ever needed to do.  Someone with a deeper understanding (like, just which algorithm the tracker software uses when you hit the "best fit" button) can be absolutely vital when you're trying to work between multiple frames of reference.