CRX-5ia Accuracy

    Hello fellow colleagues!

    I have a Granite block with certified flatness of less than 2 microns (<0.002mm) and a pointer attached to a Force/Torque sensor installed in the Robot's EoA. The force sensor is connected to a PLC that sends a digital signal to the robot whenever the force reading on Z axis reaches a threshold.

    I have setup a UF on top of that granite block using AGFS and then I made a program where the robot moves along Y axis in increments of 1mm then "attacks" the block with tool offset and high speed skip condition until it gets the signal from the PLC. Once it gets the signal it retracts to the clearance height and moves to the next position to repeat the steps. The attack motion is Linear and the speed is 1mm/s with 10% override, so, very slowly, to minimize both strain on the force sensor and joints and also to get a more accurate position recording.

    So I did this on a straight line in Y, over 17mm and 13 times in a row, collected all the PR's containing the positions registered with the high-speed skip condition and plotted Z vs Y. I was expecting an almost flat line (+/- 30 microns between the points) with a very small slope, that would indicate the angle mismatch between the granite block and the taught UF, but to my surprise, the variation between heights can sometimes reach 150+microns! The graph varies almost like a senoidal shape with peaks and valleys occurring right next to each other.

    Since I went over the same points 13 times I could notice that the height at the same spot would vary about 30 microns (which is the rated repeatability of the CRX-5). So seems like the robot is precise, but not accurate.

    The grouped dots in the red box show us the robot repeatability, varying about 30 microns at the same spot. The height difference between the points shows us the accuracy, which in this case is about 160 microns.

    Have you guys dealt with this kind of situation before?

    Thanks once again for helping out!

    Edited once, last by AlexMac (October 30, 2024 at 4:39 AM).

    Quote from AlexMac

    .. the height at the same spot would vary about 30 microns (which is the rated repeatability of the CRX-5). So seems like the robot is precise, but not accurate...

    Look for the rated precision in manual 😉.

    If you find one it will be much higher than the repeatability. If you don't find one, now you know why you don't find one. So

    Quote from AlexMac

    .. but to my surprise, the variation between heights can sometimes reach 150+microns!..

    Isn't a surprise for most people here in the forum.

    Quote from AlexMac

    Have you guys dealt with this kind of situation before?

    We deal with this fact every day. It's a common behavior of every robot.

    Hello hermann ! Thanks for the reply! So it is correct to assume that something similar is also happening with X and Y? Even though the encoder position is showing that is is the right coordinate it can vary that much?

    If I work with recorded PR instead of offsetting, does it make any difference?

    Do you know any robot manufacturer that rates accuracy? FANUC has a procedure called 'Accuracy Enhancement' that I have not yet tried, I might give it a shot then.

    Has anyone here tried this 'Accuracy Enhancement' procedure before?

    Edited once, last by AlexMac (October 27, 2024 at 6:09 PM).

    Quote from AlexMac

    Since I went over the same points 13 times I could notice that the height at the same spot would vary about 30 microns (which is the rated repeatability of the CRX-5). So seems like the robot is precise, but not accurate. On Monday I can show you guys the graph so you can have a better understanding...

    I don't agree with "robot is precise", I would agree with the statement "robot is repeatable, but not accurate".

    Its not uncommon for some to assume a robot's repeatable specification is also its accuracy - not true and from experience, definitely not. Frequently I have to point this out to mechanical designers.

    Yes, for any articulated-arm robot, Precision/repeatability is much better than accuracy. One complicating factor for robots is that "repeatability" isn't just "going to the same end point," but also includes "starting from the same start point." Driving to the same endpoint from multiple different start locations and/or orientations gives results that correspond with the robot's accuracy rather than its repeatability.

    I've just posted a picture of the graph, right in the first post....

    What I plan on doing next week or so is following the FANUC's Accuracy Enhancement document which consist basically of remastering joints J2 to J5 by jogging the robot to a pointer and changing it's configuration.

    Once I've done this I will run the same test and post the results here.

    I have notice this before, that teaching a frame or even a point in one configuration then changing it to a different one resulted in the robot being slightly off target, even though it showed being at the very same position.

    We also have a LR-Mate 200iD/14L which I plan on running this same test this week, I will post the results here just for comparison.

    Sounds like your error is based on the communication speed between the PLC and the robot. Your error looks cyclic and is not precise because the quotient of the communication speed with the PLC and the frequency at which you are doing your test is greater than 1 (not aliasing).

    If you are using the PLC for the signal and you are trying to obtain precision < .030mm, then you likely need to move much slower than 1mm/second.

    The fastest way to acquire a skip signal using force is to use a Fanuc loadcell that is connected directly to the controller. If this is not an option then try decreasing the request packet interval between the transducer/ PLC, and PLC / Robot.

    Depending on the PLC you're using, you may want to stick this logic in a high priority routine that scans much more frequently than the standard logic.

    Also, Fanuc has the ability to use highspeed digital IO. If the transducer is programmable and has a hardwired digital output, bring this signal directly to the robot.

    In my opinion, your best solution is to use a touch probe such as a Renishaw touch probe and connect it directly to the robot.

    Thanks for your input EndAffector! However I don't really think that it is a problem of undersampling an analog signal. My PLC, which is only doing this interface between my f/t sensor and my robot, has an average cycle time of 300 micro seconds; I've tried running a background logic to capture the robot's position and comparing with the High-Speed Skip and turns out it's the same, so it is safe to assume that the robot is checking its inputs every 8ms. So from the PLC to process the reading and activating is output to the robot detecting the signal and registering its position in the PR should take less than 10ms.

    I am moving at 1mm/s with 10% override, so it's about 100 microns/s. So we have plenty of time to sample and process the signal. I also tried running the entire process at 1% override speed and it does not have any affect on the results. So I am pretty sure that it is an accuracy problem from the robot side.

    But thanks a lot for your input, great way to think outside the box.

    There is ways to reach accuracy almost as good as the repeatability.

    Many robot suppliers provide an Absolute Accuracy option. These specific robot individuals have then been measured and identified at factory. Not an option you can turn on after shipment. Usually the robot supplier don't have any solution to perform it on-site. You have to de-comission the robot and send it back for recalibration. This is also the case when Absolute Accuracy robot have been repaired (switched servo or reducer) or have been involved in a collision affecting the mechanical integrity of the robot arm.

    I am working on a company that provide system and services where these kind of measurement on-site and in production cell and even with robot missing the Absolute Accuracy option from the beginning. The calibrated model take both kinematics and elasticity into account to enable off-line generated program to work without touch up.

    Quote from mr.magnuso

    Usually the robot supplier don't have any solution to perform it on-site.

    Some do. In some circumstances, tuning the robot in the working location with the actual working payload is the only way to achieve sufficient accuracy. Also, by reducing the calibrated volume to the volume where it's needed, instead of the entire robot reach volume (which is what the factory calibration does).

    Quote from mr.magnuso

    Many robot suppliers provide an Absolute Accuracy option. These specific robot individuals have then been measured and identified at factory.

    Which usually does not work very well for payloads different from the default payload.

    Frankly, I'm not sure I've ever seen factory Absolute Accuracy that was worth the money. I've had better luck with field calibrations or compensation tables, or local landmark guidance when that's possible.

    Quote from SkyeFire

    Some do. In some circumstances, tuning the robot in the working location with the actual working payload is the only way to achieve sufficient accuracy. Also, by reducing the calibrated volume to the volume where it's needed, instead of the entire robot reach volume (which is what the factory calibration does).

    It is true that the measurement is done in the full workspace using laser tracker equipment but it also performed in the full payload range. We have had the opportunity to evaluate the AbsAcc option on several ABB robots. They are very good 0.3-0.5 mm accuracy in the full workspace. We usually test them with minimum and maximum payload. With our methods, we usually could push the accuracy even further including also orthogonal elasticity (how bearings is affected by the load of robot itself and payload). ABB is what I Know never performing such measurements at customer site. So the big drawback is the risk of invalidated model is something happen to the robot (repair or other incidents). We use a method to be able to generate robot programs in the subset of the workspace as you describe, to re-create the model if needed. Since we are in the production cell, we can also in the same time probe robot to fixture or workpiece. As always, only an accurate robot is not enough, tool and workcell calibration have to be in place to be able to run off-line generated programs without touch up.

    Quote from SkyeFire

    Which usually does not work very well for payloads different from the default payload.

    Frankly, I'm not sure I've ever seen factory Absolute Accuracy that was worth the money. I've had better luck with field calibrations or compensation tables, or local landmark guidance when that's possible.

    The complexity of the system could easily increase if you need sensors. To be able to run "blind" as I describe will require also high accuracy on your products (to manipulate or process). The automotive industry have succeeded to utilize 99% off-line generated robot programs without touch up (at least company that I know about). This is not possible without AbsAcc robots. The question is if this is the way forward for other industries?

    Quote from mr.magnuso

    The automotive industry have succeeded to utilize 99% off-line generated robot programs without touch up (at least company that I know about). This is not possible without AbsAcc robots.

    Eh? My colleagues and I achieved that in automotive before AbsAcc was even an option. It simply required synchronizing mechanical design, simulation, and installation calibration. Which was more work for the departments involved, who would rather just "let the robot programmers fix everything", so adoption was spotty. But it was entirely possible.

    But automotive is easy street. The real problems arise in industries whose products require high accuracy but have very little repeatability. Aerospace, in my case, and aerostructures in particular.

Advertising from our partners