Robot Accuracy

Considerations to make when using robots in high accuracy

applications and programming robots off-line.

Image courtesy of GKN and StEM project

E XE CUTIVE SUMM ARY

This paper gives an overview of the niche field of high accuracy robotics, particularly in the
area of moving from virtual world (3D CAD robot simulations) to production where the tolerance
requirements are less than 1mm.
We explain the construction and reference systems for the robot and its environment, and the
main sources of error between nominal and reality. This includes practical examples of robot
measurements.
In summary of sources of error we show an Absolute Accuracy Robot specified as being
accurate to within 1mm of nominal. Different manufacturers have different tolerance limits around
this, and it also depends on robot size (where larger robots are generally less accurate). To help put
this in context, here is an extract from ABB Robotics IRB6640 range of robots, The difference
between a virtual robot and a real robot can be typically 8-15 mm, resulting from mechanical
tolerances and deflection in the robot structure. The Absolute Accuracy concept bridges this gap with
a complete accuracy concept for the entire robot lifetime, ensuring a maintainable accuracy of
approx 0.5 mm in the entire working range.
(https://library.e.abb.com/public/e69d8dd25cd7d36bc125794400374679/AbsAccPR10072EN_R6.pd ). We
verified this specification was correct (Average 0.5mm; 97% within 1mm; max 1.2mm). We also
measured another new standard robot measured to similar tolerance. In this case it was a standard
Fanuc M710i without absolute accuracy option but smaller/lower payload (max error approx.
1mm, average error 0.4mm improved to average error 0.3mm with calibration).

Robot Accuracy

Page 2 of 36

We share the effect of temperature (running a robot hard) increasing change in position of
around 1mm (some robot vendors offering motor cooling systems to compensate for this). We show
the effect of putting a robot on a rail causing a step degradation of several millimeters. There is data
on applying forces to robot moving it generally between 1->2mm, and when using these forces the
potential to skid and the use of metrology to understand and tune these effects although the
behavior under load is non-linear and therefore difficult to quantify certainly over time as the
machine wears.
The difference in performance between static position accuracy (we call destination) and
path accuracy (we call journey) are explained.
After explaining all of these challenges when trying to use robots in high accuracy
applications, we present some solutions. On the research side a brief summary of our work on the
COMET project in particular with Delcam and Lund University is given, for both off-line
compensation and on-line compensation for machining tasks. On the commercial side we explain
Adaptive Robot Control (ARC) where we control the position of a robot to 0.1mm, with final drilled
hole positions within 0.2mm (its not just the robot thats involved in the process tolerance budget).
We also explain our solution for robotic trimming with final cut path accuracy <0.5mm. [Note: these
are max errors 99.7% (+/-3) not average errors as some people quote].
To end we introduce a relatively new theme of multi-sensor systems to perform high
accuracy robotic production where sensors measure as built condition, and this information is
used intelligently to adapt the nominal 3D CAD process to produce quality parts and assemblies.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 3 of 36

Robot Accuracy

P ERS PE CT IVE
The current global market for robotics is approximately 100,000 units sold per year. Within
that there are some different definitions of what a robot really is but the generally accepted
version is shown in figure 1 below

Figure 1 : Industrial Robot in various deployment configurations

Figure 2. Robot Processes

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 4 of 36

Robot Accuracy

As a production tool, robots go back in time to the 1970s before the widespread use of 3D
CAD in manufacturing engineering. They replaced manual, repetitive work and were programmed
on-line using a joystick and record position button. The program then just ran repeatedly.

Figure 3. On-Line Programming

There are many variables in this process including the skill of the programmer, the quality of
the parts used to program with, and the capability of the machine. To introduce a new manufactured
product therefore takes time, a level of iteration/scrap and in many cases a level of on-going
maintenance (tweaks) through the lifetime of the product. That all being said, even today this is the
most common method of robot programming. The key customer drivers are price and reliability.
The 1990s saw the widespread introduction (in the automotive market) of 3D CAD for use in
manufacturing engineering. This led to one of the main threads of this paper the use of robot
simulation and the transfer of those programs from the virtual world into the real world. This opens
up the topic of Robot Accuracy. Also linked to this is the production use of robots on a large scale,
where high utilization is needed for commercial pay-back. So if a robot fails in the line, it needs to be

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Robot Accuracy

Page 5 of 36

replaced as soon as possible without extensive on-line programming. This is the other dimension to
robot accuracy in a car plant in particular.
A more recent niche development in robotics has been to replace much more expensive and
dedicated machine tools. This has moved the discussion on robot accuracy from @1mm in
automotive, to @0.2mm in aerospace.
This paper is focused on the niche high accuracy robotic market. It covers the different
processes where a robot either stops to deliver a process (for example a spot weld or to drill a hole),
and path following applications like trimming and adhesive application; the sources of error; and
some solutions. Our wish is this is an interesting read for anyone exploring this market.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 6 of 36

Robot Accuracy

DEST INAT ION OR JOUR NE Y

We have used the above title to describe two very different scenarios. One where we are
interested in the accuracy of the Destination where the robot performs a process for example to
drill a hole; the other where we are interested in the Journey for example laying a path of
adhesive

Destination
There are many sources of error in robotics, and these will be described fully in this section.
But the key point in this area is there is an opportunity to measure and correct before applying the
process. In many high-accuracy applications robot cycle-time isnt the factory bottleneck.

Figure 4. Transferring robot programs from Virtual World.

When transferring robot programs from virtual to the real world there are number of reference
systems used. Here is a brief description of the key elements.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 7 of 36

Robot Accuracy

Figure 5. Product Reference or Base Frame (Image courtesy of GKN and StEM project)
The first key 6D Coordinate system is from the 3D CAD of the product to be made. This
reference is used to build and certify the production jig to. So its a common reference from the part
and fixture. For the robot it defines where the part/fixture is in space, away from the machine. A
neutral name used in many robot simulation systems is called the Base Frame but many robot
venders have a unique name. This reference system is the primary reference for the part to be
produced.

Figure 6. Target (Image courtesy of GKN and StEM project)

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 8 of 36

Robot Accuracy

The next reference frame is the Target (Figure 6) where the process takes place. In this
example it is where the hole is drilled. The robot simulation system makes sure this is defined normal
(90) to the surface of the original 3D CAD design. So in the virtual world we have a perfect
process. All of the targets are 6D frames relative to the product reference frame (Base Frame in
Figure 5).

Figure 7 Robot Reference Frame (Image courtesy of GKN and StEM project)
The next reference system is for the robot system. As with many such references in a virtual
world they may be in space not a physical reference. In the case on Figure 7 the robot is mounted
on a rail, and the robots reference system begins at the origin of that rail system. When transferring
from virtual to real worlds the position of the Base Frame (Fig 5) from the Robot Reference Frame
(Fig 7) is a major potential source of error, outside of the accuracy of the robot arm.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 9 of 36

Robot Accuracy

Figure 8. Robot World Frame (Image courtesy of GKN and StEM project)
The Robot when delivered out of the box, has its own reference system at the base of its
kinematic chain. This is normally at the base of the robot where it mounts to the floor (or in our case
the robot rail) and concentric with its first rotary axis (axis 1). Whilst this isnt always the case, more
often than not it is so useful for this initial description.

Figure 9 Tool Frames (Image courtesy of GKN and StEM project)

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 10 of 36

Robot Accuracy

A standard industrial robot will have 6 rotary joints, describing a kinematic chain from its base
through to the mounting flange where the tool is attached to. The naked robot out of the box has a
tool flange with a frame often called Tool Zero (Tool 0) because nothing is attached. Then the
process tool (in this example a drill) is attached and a unique tool is defined in the robot controller.
This is usually a 6D Frame at the point of the tool which you want to drive to the Target so in this
example the tip of the drill to go to the place described from the original 3D CAD design to make a
hole. A payload (weight of tool) and Centre of Gravity are also required in the robot controller for
the motion to be controlled with the correct inertial loads. So the defined tool (sometimes called TCP
for Tool Centre Point) is the end of the robotic kinematic chain.

Figure 10. Linking Robot to Fixture (Image courtesy of GKN and StEM project)
To link the robot system to the fixture a Base Frame is defined which is where the fixture is
with respect to the robot system.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 11 of 36

Robot Accuracy

Figure 11 Full Kinematic chain (Image courtesy of GKN and StEM project)

Figure 12. Inverse and Forward Kinematics

Knowing where the robot is (current position of TCP), and where it needs to go (Target) the
robot controller computes how to get there (inverse kinematics) it plans its motion (forward
kinematics) and tracks progress through encoder feedback to get there. In the robot simulation

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 12 of 36

Robot Accuracy

environment the target is reached but on the shop floor with sources of error it gets close (to a
tolerance). What many people ask is, How Close?

Figure 13. Kinematic errors

The first source of error we will discuss is called Kinematic Error these are errors in the
construction of the robot (from nominal). At a simple level it could be the incorrect set-up of the robot
(called mastering) where there may be an error in encoder definition (where zero may not be
actually zero degrees for example). It could be the distances between the joints (link lengths) or
angular errors in construction between the joints. As this is a chain small errors at the start of the
chain can produce large errors at the end.

Figure 14. Measuring equipment

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 13 of 36

Robot Accuracy

To measure/investigate we use Nikon Optical CMM called K-Series. This can track multiple
6D frames simultaneously for a full understanding of the robot behavior. Measurements can be
dynamic as well as static, to give a full understanding of the machine. For measurements of robot for
Kinematic Errors we use Nikon software (ROCAL which stands for RObot CALibration)
For measurement on a naked robot (as an example) we measure the base of the robot, and
the robot flange (tool 0). We instrument the robot flange to track the robot in 6D, and create a robot
program of 50 random points to go to. In this example, its a new, 50Kg payload robot with four
points we simultaneously track on a bracket attached to the robot flange. The reporting is the
difference in position between nominal and reality.
Residual Position Error in markers 1 to 4
1.1
Epos1
Epos2

Epos3
0.9

Epos4

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

10

15

20

25

30

35

40

45

50

Point Number

Figure 15. New robot measurement

It can be seen that a couple of positions are just over 1mm out from nominal, and the
average accuracy is 0.37mm. This is absolutely normal. By exploring the sources of error we can
improve the results. In this case by looking at encoder errors the maximum and average errors are
reduced. In this case just minor corrections to joints that should be at zero (J2: -0.0187
J3: 0.0162 J4: -0.0148 J5: -0.026691) gives an average error of 0.278mm. Note Joint 1 at the start
of the chain would be covered by Base Frame and Joint 6 at the end, by the Tool Frame so not
included in this adjustment.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 14 of 36

Robot Accuracy

Figure 16. Results adjusted for encoder offsets.

These results are then applied to the robot controller and a new 20 random point
verification program created and measured to confirm the result. There are two methods to apply this
knowledge one is to update the robot controller; the other is to filter the robot program off-line to
correct for these errors (if access cannot be made to controller parameters).
To bring into context, small robots are more accurate (generally) than large because the
amplification of errors is less. Also they have a smaller resolution / incremental step capability to
adjust to correct position. For most robot users the numbers are fine; its just the customers looking
for high accuracy are sometimes surprised that the accuracy may be 1mm+ which is ten times larger
than repeatability.

Figure 17. Subset of robots measured past 15 years

The above figure shows a number of robots we measured for Airbus in the mid 2000s and
we continue to perform measurements for robot manufacturers and their clients today. We have
never measured a new robot with a payload over 5Kg that is accurate to an aerospace drilling
tolerance of 0.2mm. In aerospace, the CNRC in Montreal investigated doing predictive corrections to

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 15 of 36

Robot Accuracy

robot programs off-line, based on calibration data with maximum success of 90% (meaning 10% of
all holes out of tolerance).
In the automotive industry metrology systems are used to create an Absolute Accuracy
model / file of the individual robot, at the end of its production to facilitate robot exchange to a
tolerance of 1mm. This works well in many cases, but robots do wear out over time and if the
program has been manually adjusted to compensate a new robot may not go to that old position to
these tolerances. As such we do measure production cells for archiving purposes to adjust
programs off-line. One point of caution is if there is a calibration filter on the robot it may conflict
with another calibration filter used off-line. Its always advisable to check, before an emergency
arises.

Figure 18. Backlash / lost motion.

New robots, straight from production tend to be tight meaning little play in the gearboxes.
Most robots have motor/encoder on the same side (input) to gearbox; not on the output. This is
because for most users thats fine for performance and cost effectiveness. But as robots wear, and
get some slop changing directions may lead to lost motion (where the output doesnt move as the
gears move from one side of mesh to the other). This can be measured by driving to a point from
one direction; and to the same point from another direction measuring the difference. Some
integrators do put dual encoders on robots to compensate for backlash but generally the
intervention is a repair overhaul, required when backlash becomes significant. As errors are
amplified through the kinematic chain; axis 1 is usually the most significant and why some
integrators mount robots on 10 wedges to pre-load the gearbox. Axis 2 and 3 are pre-loaded by
gravity so become more significant if changing directions at high speed. Backlash can confuse
robot calibration tasks so best to measure separately, especially on older robots.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 16 of 36

Robot Accuracy

Figure 19. Robots and temperature

Its good practice to calibrate and run robots at their normal duty cycle (work rate). This is
because as they get hot they expand along the serial chain - so end up in a different place! Its not
unusual in an automotive plant to let the robots do some Tai Chi before shift starts (to warm-up). We
ran a series of robots from cold: very hard (axis 1, 2, 3) in 30 minute periods; 3 times. In all cases the
robot position had shifted around 1mm (between 0.9mm and 1.4mm)

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 17 of 36

Robot Accuracy

Scale
We made some tests at a UK Research Centre (AMRC) to test the out-of-box accuracy of
robot when its reach is extended by mounting it on a track (7th axis). [Presented at SAE 10AMF-077,
2010].

Figure 20. Sample data of a robot on the floor

We started with the robot on the floor, with a traditional 50 point ROCAL measurement. The
robot was an Absolute Accuracy Robot specified as having 97% of points within 1mm from
nominal, max 1.2mm.
(https://library.e.abb.com/public/e69d8dd25cd7d36bc125794400374679/AbsAccPR10072EN_R6.pd )
At this confidence limit the measurement proved it was in specification (although there were
outliers over 1mm, as expected in the specification).
We then mounted the robot on a track (7th axis) and moved to a grid of 25 points, from three
different positions on the track.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 18 of 36

Robot Accuracy

Figure 21 impact on accuracy when increasing scale

Figure 22. Accuracy of robot on slide.

As stated earlier the first issue / source of error is defining where the origin of the robot
system is to start with, and then errors to jig (Base Frame) and tool (Tool Frame) are amplified
because of the much larger distances. So the results were not a surprise. They are a combination of
error amplification in definition of the Base Frame in all positions, and then robot arm errors on top
where it is clear track position 2 has the largest reach on the arm.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 19 of 36

Robot Accuracy

Figure 23. Perceptual Map of robot accuracy vs working volume.

We conclude our kinematic error / accuracy discussion with a summary of our findings. From
Figure 23, a small (low payload robot) is more accurate than a larger, heavier payload robot. But it
has a smaller working volume. To increase working volume the robot is mounted onto an external
rail (called a 7th axis). Of course the rail can be mounted on another rail, which is then mounted on
another rail, Adding more additional axis to the robot system, and increasing working volume. The
downside to this is we see a step change in accuracy. So a small robot on a rail, would tend to be
less accurate than a larger robot mounted to the floor, covering the same volume. Once on a rail the
accuracy degrades over distance, in the same way as accuracy is difficult to maintain on a floor
mounted robot using more and more of its available workspace. So, typically we would look for a
small payload robot on a track to give the best performance. From our experience at AMRC, a robot
working comfortably in its working range, on a longer rail; is spatially more accurate than the same
robot having to stretch to cover the same work volume, on a smaller rail.

From all our work with Airbus (and others that followed) was its not possible to
move robots from the virtual world to shop floor to high accuracies (0.2mm) using
out-of-box robots, calibration and predictive algorithms.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 20 of 36

Robot Accuracy

W h a t h a p p e n s w h e n y o u a r r i v e ( P ro c e s s )

Figure 24. Force and Robots

In the same way we see lost motion (backlash) when driving the motor and changing
directions. Once stopped (arrived at destination ready to do work) adding a force into robot structure
will move it without being seen by the encoder on the input side of the joint.

Figure 25. Applying pressure to robot

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 21 of 36

Robot Accuracy

We tested this effect for Airbus, because their baseline drilling solution is to apply a massive
force to the structure (pressure foot) to be larger than axial forces in drilling, and to push stacks
together to prevent inter-laminar burring. In all cases the deflection was very large (1-2mm for robots
with payload capable of handling Airbus development drill), and indeterminate (not predictable
because it was random around robot working envelope depending on robot configuration / position).
We did instrument up each axis, and saw no deflection in the links; all the movement can be
described by the joints. At the time of writing we have yet to measure a dual encoder robot but
without a dual motor to pre-load the gearbox in both directions one assumes there is still lost motion
but by measuring the output the motor responds which could end-up competing against applied
pressure. That would be influenced by latency in motor response

Figure 26. Tuning clamp-up

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Robot Accuracy

Page 22 of 36

We have supplied services to measure the effect of high clamp-up forces which can lead to
tool skidding to a wrong position, before hole drilled. In this case the pneumatics were adjusted to be
slower / less aggressive to stop the skid but as mentioned before the effect of this was not uniform
across the robot work envelope.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 23 of 36

Robot Accuracy

Journey
The Journey describes a continuous movement of the robot, when the process happens in
parallel to the robot motion. Clearly in this scenario its not possible to stop and adjust the robot
whilst it is arc welding, dispensing adhesive, or machining.

Figure 27. Robot Paths

This example is based around trimming a window aperture in an aerospace fuselage but
equally applicable for any path application

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 24 of 36

Robot Accuracy

.
Figure 28. Robot Paths Measurements
When we look at the dynamic measurements of a robot path, a low frequency oscillation is
seen. This is typically 7-10Hz as the 6 rotary axis of a robot interpolate straight lines. We have tested
the path in different positions in the robot envelope (different Base Frames) and see the same
pattern. We have tested our markers on a very accurate bridge coordinate measuring machine, and
sensor noise is 10-30m so the oscillation isnt from the metrology system.

Figure 29. Robot behaviour on paths part 2

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 25 of 36

Robot Accuracy

Looking at the measurement of the complete path we see other machine behaviours. From
Figure 29 as the robot makes a big change in direction (top right) this is where the robot is furthest
from base of the machine, and we see a combination of flexibility in the tool and some backlash /
overshoot in the robot. As it enters the first corner, and exists the second corner we see there is a
spike which is caused by the robot changing is configuration. An industrial robot can get to the
same place in a number of different ways, called solutions. Axis 4 and 6 are usually concentric (at
least nominally) with axis 5 orthogonal so its not unusual to perform a straight line, and change
configuration of joints. What is measured is the spike in making that transition.

Figure 30. Robot repeatability

But the repeatability of the path is incredible
From many projects we know there are many variables in programming robot paths. But in
terms of accuracy we see lower magnitude oscillations if the robot runs slower. As the robot
controller is looking ahead, we find programming extra points at the exits of bends reduces
overshoots. The robot controller, when doing its mathematics on forward kinematics takes the high
level program from the robot simulation system, and does its own path planning from that. So if its a
straight line, we see no benefit in adding additional points along that line (because mathematically
its what the robot is trying to calculate motion on anyway. However, if those points are a reference
to be moved due to some calibration then indeed extra points are useful (because the adjusted
path will not be calculating the same straight line.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 26 of 36

Robot Accuracy

S OLUT IONS
O f f- L i n e

Figure 31. Off-line Compensation

For Off-Line compensation the goal is to measure the sources of robot error, and apply them
to the nominal 3D robot simulation. The logic being you know where you want the robot to go (RED
path above in Fig 31), you know where it may deviate from that program (GREEN path), so apply the
opposite (BLUE path), which is actually run on the robot, with the result you achieve BLACK path
which is closer to your goal than the original GREEN path.
We have described using Nikons ROCAL software for doing this with kinematic errors, but
recently we worked on a European R&D project called COMET. DELCAM develops software for
creating robot programs off-line. They created an infrastructure for creating an individual robot
signature file.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 27 of 36

Robot Accuracy

Figure 32. COMET Robot Signature (DELCAM)

In this scenario, the kinematic errors are measured using NIKON metrology system (KSeries). But also the non-linear behavior in the joints is measured and the process forces to give a
holistic compensation.
Lund University in Sweden has an international reputation in non-linear control systems and
robotics. They have developed a new (Patent Pending) method of describing the non-linear behavior
in the joints which historically has been an issue in compensation because its non-linear
(different in different positions/configurations of the robot arm, throughout the program).
DELCAM have used their experience in 5-axis machining (POWERMILL product) to measure
and compensate for machining forces.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Robot Accuracy

Page 28 of 36

Figure 33. COMET Joint Based Model (LUND University, commercialized through its spin-off
Cognibotics)
The results to-date show promise with some results being within 0.2mm on small test
coupons (limited size) but also highlighting good practice mentioned earlier. One example being
to create the nominal simulation on one part axis 1 was locked; but when compensations were
applied there were a number small movements which introduced the effect of backlash, and the
results were worse. But these are education/method issues. The clear advantage is extra sensors
are not needed in the robot cell, and improvements from out-of-box robot accuracy shown. But the
time period between measurements is yet to be established for production use.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 29 of 36

Robot Accuracy

Re a l -T i m e
On the same COMET project the partners worked on real-time control. The benefit is clear
to have a live adjustment of the robot, to an accuracy much higher than it could achieve in a large
volume, out-of-box. The focus is on the Journey so path applications, especially machining
(because of DELCAM software/experience). As well as DELCAM, Lund University and Nikon there
were a number of integration partners / test set-ups, including SIR in Italy, Fraunhofer IPA Stuttgart
in Germany, and AMRC in UK.

Figure 34. State of Art Robot Control

The measurement system used was Nikon K-series, with High Frequency (HF) option,
measuring up to 1,000Hz for one 6D frame. The most advanced deployment was with Lund
University in Sweden. In this case there were two 6D frames (one for Tool Frame; one for Base
Frame) measuring at 500Hz, feeding into Lund software (note all software from Nikon and Lund was
running on same industrial PC). Every other measurement was then streamed into the robot
controller (running at 250Hz).
This set-up solved two key issues. The first is latency. The latency on the Nikon
measurement system wasnt noticeable in the process but from graphs/reporting estimated to be
less than 3ms. For longer robot programs though more work on clock synchronisation is needed. But
the key leap was the innovative correction.
If we take it that the robot is in the wrong place (to a tolerance) and performing a path which
is planned by the robot control system, we have a gap in information. The robot doesnt know where
it is (truly) but planning where to go next. The measurement system knows how far away the robot is
off-path but doesnt know where its going next to advise a correction (which is needed at a very
low latency)

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 30 of 36

Robot Accuracy

Figure 35. Real-Time Interface

The solution was to make the correction based on the low-level side of the robot which is
innovative. So not at the high level (robot simulation -> robot inverse/forward kinematics); but
intercepting the output to the drives / adjusting it, and sending to servos.

Figure 36. Results Real-Time Interface

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 31 of 36

Robot Accuracy

A dapt ive Ro bo t Co nt ro l

Figure 37. Adaptive Robot Control (ARC)

Nikon Metrology has a technology called Adaptive Robot Control (ARC), which we developed
and has been patented by Airbus. We have commercial exclusivity. The key features are we
measure multiple frames (which other devices like a laser tracker cannot do) at high frequency (up to
500Hz; max measurement frequency 1000Hz). The camera system needs to see the robot tool
(Tool) and instrumentation for the base (Base). But its measurement volume is flexible (camera can
be moved in a larger volume, to change area of interest), and often used to measure parts far larger
than its field of view as proven in many years of production.
So we can measure a robot tool to a part held by another robot, running through a series of
different orientations to process a part, to a very high accuracy (better than 0.1mm)
For paths we run a dry path (no corrections) and then apply the corrections on the next path /
subsequent paths. We do it this way because making corrections point-by-point may run against the
mesh of the gears (which can cause the next path to be worse) but because we always measure
live dynamically we check the path accuracy against warning limits and then stop to re-do the dry
run again. This has been proven in TRL gateways, but not in production yet. The major
benchmarking of this technology was on HondaJet for GKN Aerospace USA for composite trimming,
pocket milling and edge milling which is where the maximum error of 0.5mm comes from (mean
error less than 100m).

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 32 of 36

Robot Accuracy

Figure 38. How Adaptive Robot Control increases accuracy

The key feature in ARC is we measure the robot where it is (Tool and Base), and guide it to its
nominal position. This is independent of the entire kinematic chain.

Figure 39. Adaptive Robot Control on Robot mounted on 7th Axis Rail (ABB robot system at AMRC
described earlier

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 33 of 36

Robot Accuracy

The results using ARC on the ABB Robot/rail at AMRC showed all final robot process positions
(verified by a Laser Radar and Laser Tracker) were less than 0.2mm we specify so more than an
order of magnitude better than a calibrated robot cell and no outliers, or any significant differences
with robot reach.

Figure 40. Adaptive Robot Control with RLO

Using a special feature of ARC called RLO (Remember Last Offset); we do a dry run measuring
each point but not applying a correction as that may run against the meshing of the gears
(backlash effects). The corrections are applied to all the target points at the end of the path run, and
then a very accurate path can be run to do the process. Using ARC with RLO at GKN Aerospace
USA all measurements were within 0.5mm (3 = 0.48mm) against a customer specification of
0.75mm (30/1000).

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 34 of 36

Robot Accuracy

Figure 41. Latest ARC Benchmark 2015

The above plot is from a recent aerospace benchmarking project, with a new robot (not
absolute accuracy, but small payload), and a carefully set-up cell (robot payload, Tool Frame and
Base Frame carefully defined). We measured each robot position (defined by the Off-Line 3D CAD
robot simulation). The results are entirely within vendor expectations but often not the customers. It
highlights how spatial accuracy is very different from repeatability, and the power of the metrology
system with Adaptive Robot Control, to dramatically improve accuracy.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Page 35 of 36

Robot Accuracy

Multi-Sensor

Figure 42. Multi-Sensor Robotics

A recent theme in accurate robots has been the addition of multi-sensors inside the
production cell. The main image above shows a robot held laser scanner, which is tracked by a KSeries camera so actually completely independent of robot accuracy. This measures the outside
geometry of the as-built part, giving information to make intelligent decisions on the process
automation.
The other images are for NDT, where again the sensor (in this case an ultrasonic probe) is
tracked in 6D (independent of the robot) by the same K-Series optical CMM. This gives information
on the inside of the part being processed in a common 3D reference system, allowing direct
comparison to original 3D CAD (design intent).
The combination of the two bits of information is used to Adapt the robotic process.
Implementation depends on the accuracy and environment of the robotic process. For very high
accuracy aerospace applications, the integrated use of a live metrology technology like Adaptive
Robot Control achieves +/-0.2mm final process (with respect to the Adapted 3D CAD nominal).
Some processes require an improvement on robot accuracy beyond absolute accuracy so higher
than 1mm, but wider than 0.2mm. In these scenarios a combination of off-line calibration and
backlash compensation is a very good solution.
Offline compensation of backlash and other deviations such as non-linear compliance is
inherently difficult since the parameters depend on each individual robot, its production build, wear
in service, application, environment, and so on. One company at the leading edge of this research is
www.cognibotics.com with their patent pending technology to measure and compensate this
behaviour, validated by K-Series optical CMM.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]

Robot Accuracy

Page 36 of 36

At times the process also includes a pressure requirement, which brings in yet another
sensor technology (force feedback), and its the combination of many sensors to achieve high
accuracy robot automation.

Acknowledgements
A huge number of people, over many years have been involved in the production of this paper
(literally too many to mention). Organisations partnered; projects undertaken; knowledge acquired
all helping our community at pushing the boundaries of using standard industrial robots to make
use of their great prices (because of the economies of scale created by the volumes they make) but
in novel cost efficient, high-accuracy applications.
But special thanks is needed for some
Long lasting colleagues, Richard Kingston and Paul Lightowler
The family trees of Krypton->Metris->Nikon Metrology in this field
All our customers who have invested in our solutions
The COMET project team.

NIKON METROLOGY
Nottingham EMA, Argosy Road
Castle Donington, Derby, DE74 2SA, UK
Tel.: +44 1332 811 349 Fax: +44 1332 639881
[email protected]