[comp.robotics] High Accuracy Manipulators

barts@cyber.Eng.Sun.COM (Bart Smaalders) (08/30/90)

In article <1990Aug30.014817.8794@portia.Stanford.EDU> boehlke@sunrise.stanford.edu (Dan Boehlke) writes:

   I see the next breakthrough in robotics being the 
   introduction of very high accuracy manipulators--
   say an order of magnitude (or more) better than
   any systems of the 80's.  There are plenty of potential
   products that simply cannot be assembled today anywhere 
   outside of a laboratory.

Very high accuracy manipulators are very difficult to design using
today's technologies if they are to be of any general use.  The best
(gp commercial) figures I remember from the 80's are approx. .001"
repeatabilty for a _light duty_ electronic assembly robot (working
envelope ~ 1 cubic foot).  Building a manipulator _accurate_ (not just
repeatable) to .0001" with a similar envelope would probably imply:

  a.  All major joints must be linear slides.  Rotary motion is ruled
out if arms are of interesting lengths, since accuracy and stiffness
requirements go through the roof as arm length increases.  In addition,
kinematic errors due to inaccurate knowledge of zero position are
difficult to reduce at these accuracy levels.

  b.  Temperature controlled environment and workpieces is a must.  Any
sources of heat in the arm itself must also be cooled, or made constant
in effect.  A motor mounted in an arm will cause measurable deflections
as the motor heats up and warms the surrounding structure.

  c.  The tooling (end effectors, part fixturing, etc) costs will be
very high.  The assembly cell should be mounted on a isolated granite 
slab.

  d.  Frequent recalibration and verification would be a must - this
would be a rather delicate device.  

A interesting starting point for such a manipulator would be to look at
Coordinate Measuring Machines and the technology used to build them.

- Bart 

barts@Eng.Sun.Com
--

---------------------------------------------------------------------------------- 
 Bart Smaalders  Sun Micro Inc        | This space available 
 barts@cyber.sun.com                  | 

loucks@intvax.UUCP (Cliff Loucks) (08/31/90)

From article <141582@sun.Eng.Sun.COM>, by barts@cyber.Eng.Sun.COM (Bart Smaalders):
< 
< In article <1990Aug30.014817.8794@portia.Stanford.EDU> boehlke@sunrise.stanford.edu (Dan Boehlke) writes:
< 
<    I see the next breakthrough in robotics being the 
<    introduction of very high accuracy manipulators--
< 
< Very high accuracy manipulators are very difficult to design using
< today's technologies if they are to be of any general use.  The best
< (gp commercial) figures I remember from the 80's are approx. .001"
< repeatabilty for a _light duty_ electronic assembly robot (working
< envelope ~ 1 cubic foot).  Building a manipulator _accurate_ (not just
< repeatable) to .0001" with a similar envelope would probably imply:
< 
< 
< A interesting starting point for such a manipulator would be to look at
< Coordinate Measuring Machines and the technology used to build them.
< 
< - Bart 

That's essentially what Adept did with their new UltraOne.  It's a
four axis gantry on a granite base.  Rated accuracy is +-0.0002" in
a working volume of 26" x 26" x 8".

Cliff
-- 
A society is not civilized until it domesticates the icecube.

Cliff Loucks  <=>  loucks@intvax.UUCP
Sandia National Labs, Albuquerque, New Mexico

mgsmith@hplabsb.HP.COM (Michael Smith) (08/31/90)

>In article <1990Aug30.014817.8794@portia.Stanford.EDU> boehlke@sunrise.stanford.edu (Dan Boehlke) writes:
>
>   I see the next breakthrough in robotics being the 
>   introduction of very high accuracy manipulators--
>   say an order of magnitude (or more) better than
>   any systems of the 80's.  There are plenty of potential
>   products that simply cannot be assembled today anywhere 
>   outside of a laboratory.
>
>Very high accuracy manipulators are very difficult to design using
>today's technologies if they are to be of any general use.  The best
>(gp commercial) figures I remember from the 80's are approx. .001"
>repeatabilty for a _light duty_ electronic assembly robot (working
>envelope ~ 1 cubic foot).  

One reason humans are capable of high accuracy tasks (putting a chip
on a circuit board for example) using a low accuracy manipulator
(the human arm) is because of the use of end point control.  One looks
at what one is doing while doing it.  Instead of making a robot more accurate, 
the loop needs to be closed at the end point using either vision or other
sensors.  Then only the resolution of the robot is important which of course
is much simpler than improving the accuracy of a manipulator.

End point control is also advantageous because it can eliminate the 
need for special linear slide joints, a temperature controlled environment, 
an isolating granite slab, and frequent recalibration.

Mike Smith
HP Labs

barts@cyber.Eng.Sun.COM (Bart Smaalders) (09/01/90)

In article <5829@hplabsb.HP.COM> mgsmith@hplabsb.HP.COM (Michael Smith) writes:

   One reason humans are capable of high accuracy tasks (putting a chip
   on a circuit board for example) using a low accuracy manipulator
   (the human arm) is because of the use of end point control.  One looks
   at what one is doing while doing it.  Instead of making a robot more accurate, 
   the loop needs to be closed at the end point using either vision or other
   sensors.  Then only the resolution of the robot is important which of course
   is much simpler than improving the accuracy of a manipulator.

   End point control is also advantageous because it can eliminate the 
   need for special linear slide joints, a temperature controlled environment, 
   an isolating granite slab, and frequent recalibration.


True to a point, especially with more conventional accuracy requirements-
but the discussion was about manipulators with ~ .0001" accuracy.  This
would imply some very accurate sensors indeed.  Unless the sensor system
directly measures the positioning error (eg end-effector to workpiece) as a
differential error , one would still need to calibrate the sensor to the
end-effector. Most vision-based sensor systems would have considerable
trouble at this resolution level, and would have a very small field of view.
A contact-based sensing systems is generally much more accurate, but also is
more delicate, has more "observability" problems and recalibration is more
difficult.

In addition, for most assembly tasks the sensor-based system would be
considerably slower, since the robot would need to come to a stop (with time
to damp out all vibration to well below .0001") at least twice before
attaining the final position.

The original poster was interested in moving this sort of technology out of
the lab...that isolated granite table is still probably required, since the
forklifts running down the corridor outside tend to make the robot shake a
thou or two....  And I wonder how stable the camera and lens are over the 20
degree temp variation we get between the night and day shifts :-).

I'm still not sure what product really requires this level of accuracy in
assembly - can anyone clue me in?  It would be interesting to examine the
error budget for the complete system.

- Bart

barts@Eng.Sun.Com

--

---------------------------------------------------------------------------------- 
 Bart Smaalders  Sun Micro Inc        | This space available 
 barts@cyber.sun.com                  | 

ulrich@grip.cis.upenn.edu (Nathan Ulrich) (09/02/90)

An earlier posting says (sorry I don't have the reference):
   I see the next breakthrough in robotics being the 
   introduction of very high accuracy manipulators--
   say an order of magnitude (or more) better than
   any systems of the 80's.  There are plenty of potential
   products that simply cannot be assembled today anywhere 
   outside of a laboratory.

Dan Boehlke (boehlke@sunrise.stanford.edu) replies:
   Very high accuracy manipulators are very difficult to design using
   today's technologies if they are to be of any general use.  The best
   (gp commercial) figures I remember from the 80's are approx. .001"
   repeatabilty for a _light duty_ electronic assembly robot (working
   envelope ~ 1 cubic foot).  

Michael Smith (mgsmith@hplabsb.UUCP) responds:
   One reason humans are capable of high accuracy tasks (putting a chip
   on a circuit board for example) using a low accuracy manipulator
   (the human arm) is because of the use of end point control.  One looks
   at what one is doing while doing it.  Instead of making a robot more accurate, 
   the loop needs to be closed at the end point using either vision or other
   sensors.  Then only the resolution of the robot is important which of course
   is much simpler than improving the accuracy of a manipulator.

High precision manufacturing can be handled in two distinct ways.  The current
automated approach is to use very massive and very stiff machines with linear
slides and precise position control.  There are grinding operations which
are toleranced in millionths of an inch (that's 0.000001") and the machines
that can handle this are vibration-isolated and temperature controlled.  

You will never get near this precision with a serial robot arm--the geometry of
cantilevered links connected by revolute joints just does not lend itself
to this kind of accuracy.  Some small SCARA arms claim repeatability under
controlled conditions of 0.001" inch, but this is mostly BS.  I challenge
anyone to show me a serial robot arm that can obtain accuracy (not
repeatability) of better than 0.01" inch outside of controlled testing and
calibrating environments, and throughout its claimed payload range.

But we've all heard about engravers that could put the Gettysburg Address on
the head of a pin by hand.  How is this possible, given that the human arm
has lousy position control in free space, even in comparison to robot arms?
You may claim that the engraver uses a magnifying device and his advanced 
vision system to perform "endpoint" control, and if robot vision systems were 
as good we could do the same thing with robot manipulators.  Ridiculous!

The engraver does the same thing that we do when we write:  he grounds his
high workspace, low accuracy arm to the workpiece and uses his small workspace
high accuracy hand to perform the work.  Have you ever tried to write anything
small and legible without resting your hand on the paper?  "Endpoint" control
will not overcome the accuracy limitations of a robot arm, which are related
to position sensing, stiction, and stiffness.  And any incidental vibration
can destabilize such as system--not a problem if the workpiece and the
manipulator are in contact.

But I'm not a pessimist.  I think high-accuracy operations can be realized
with current technology and with serial robot manipulators.  But not with
precise position control and not with "endpoint" control, but with force
control.  This same low-accuracy human manipulation system can locate its
two hands relative to each other in space with enough accuracy to put a
0.9995" peg in a 1.0000" hole.  How?  By using compliance and force control.
This is better than any robot manipulator can accomplish without huge chamfers
and the use of a RCC (which is passive compliance).

My opinion only.  Now tell me why I'm wrong.

Nathan Ulrich                           "If it was easy, someone would have
ulrich@grip.cis.upenn.edu                done it already..."
DoD #0080 - GT750 pilot

mgsmith@hplabsb.HP.COM (Michael Smith) (09/04/90)

In article <141710@sun.Eng.Sun.COM> barts@cyber.Eng.Sun.COM (Bart Smaalders) writes:
>
>True to a point, especially with more conventional accuracy requirements-
>but the discussion was about manipulators with ~ .0001" accuracy.  
 ...
>Most vision-based sensor systems would have considerable
>trouble at this resolution level, and would have a very small field of view.

A system with an accuracy of ~2.5 microns (0.0001") that uses 
machine vision currently exists.  A small field of view is not very 
important because the features of parts that require that degree of 
accuracy are usually also very small.  After all, everything is relative.

>
>In addition, for most assembly tasks the sensor-based system would be
>considerably slower, since the robot would need to come to a stop (with time
>to damp out all vibration to well below .0001") at least twice before
>attaining the final position.
>

The coordinate measurement machine called the Ultra 1 (from Adept) is 
actually pretty slow because it is always accurate, even it doesn't
need to be.  For example, when you pick up a small part from a 
feeder, you don't usually need to know its location very precisely.
The Ultra 1 still has to carry its huge granite links around though
which slows things down.

>The original poster was interested in moving this sort of technology out of
>the lab...that isolated granite table is still probably required, since the
>forklifts running down the corridor outside tend to make the robot shake a
>thou or two....  And I wonder how stable the camera and lens are over the 20
>degree temp variation we get between the night and day shifts :-).
>

Assembling parts to ~2.5 microns can be very difficult just as producing
chips with fine traces is.  You wouldn't produce chips on the factory
floor because of course a clean room is required.  Similarly, a special
environment is usually required for high precision assembly.  The robot
system should also be mechanically isolated from the floor using
commercially available dampeners.  

>I'm still not sure what product really requires this level of accuracy in
>assembly - can anyone clue me in?  

My guess is that anything that requires ~2.5 microns accuracy is highly
proprietary so private companies (such as HP) will be very secretive.
Electronic devices in general are shrinking so there should be very
many applications in the future.

Mike Smith
HP Labs

mgsmith@hplabsb.HP.COM (Michael Smith) (09/04/90)

In article <29067@netnews.upenn.edu> ulrich@grip.cis.upenn.edu (Nathan Ulrich) writes:
>
>But I'm not a pessimist.  I think high-accuracy operations can be realized
>with current technology and with serial robot manipulators.  But not with
>precise position control and not with "endpoint" control, but with force
>control.  This same low-accuracy human manipulation system can locate its
>two hands relative to each other in space with enough accuracy to put a
>0.9995" peg in a 1.0000" hole.  How?  By using compliance and force control.
>This is better than any robot manipulator can accomplish without huge chamfers
>and the use of a RCC (which is passive compliance).
>
>My opinion only.  Now tell me why I'm wrong.
>

I work for a $10 billion a year company and I have never seen
an application where such a peg is put into a hole!  Such an application
is about as realistic as bin picking or stacking legos.  It doesn't matter
how well a system does something that is not useful.  This is a major
contention of mine.  I believe that a great deal of robotics research
is misdirected.  Hey, sure such stuff is fun.  My first paper was on
bin picking and I have given demonstrations on compliance using a force
sensor because it was  technically challenging and interesting but
subsequently I found that it was misguided.  I have never seen force 
compliance used in an actual application, not because it is unavailable,
but because it is not as needed as other technologies.

We have been collectively doing research in robotics for many years now
and they are still not widely used in manufacturing.  Why?  Because they
are not cost effective and because they have been frequently misapplied.
We need to really look at what robots should be used for and then work
on the problems that specifically prevent them from being used in
those applications.

Mike "my favorite project was a robotic bartender" Smith
HP Labs

ulrich@grip.cis.upenn.edu (Nathan Ulrich) (09/06/90)

In article <29067@netnews.upenn.edu> I write:
    ...This same low-accuracy human manipulation system can locate its
    two hands relative to each other in space with enough accuracy to put a
    0.9995" peg in a 1.0000" hole....

In article <5832@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) responds:
    I work for a $10 billion a year company and I have never seen
    an application where such a peg is put into a hole!  Such an application
    is about as realistic as bin picking or stacking legos.  It doesn't matter
    how well a system does something that is not useful.

You must not have had any exposure to mechanical systems, then, or ever spent
*any* time watching a machinist (the size of your company doesn't lend
any credibility to your ignorance, by the way).  This type of task, which has
been generalized into the peg and hole insertion example, is very prevalent
in all types of mechanical assembly.  Shafts and bearings frequently have this
type of tolerance (or tighter).  Bearing housings and bearings commonly have
interference fits (where the housing bore is *smaller* than the OD of the
bearing).  Even the relatively simple task of inserting a screw or bolt into
a clearance hole is an example of peg and hole insertion, although usually
with much lower tolerances.  I could spend pages listing the tasks I'm aware
of that fall into this category (and I'm only a PhD student, not an employee
of a multi-million dollar company :-).

I think one reason the peg and hole insertion has been a prevalent watershed
for robotic assembly is because it is a good measure of either the precision
of the system or its ability to accomplish difficult tasks with force
control or compliance.  And the tolerance of the fit is a good measuring
stick.  Systems that can accomplish tight tolerance peg and hole insertion
are also considered able to handle different geometries (square key in
square hole, etc) although this is not always true.

Michael Smith continues:
>We have been collectively doing research in robotics for many years now
>and they are still not widely used in manufacturing.  Why?  Because they
>are not cost effective and because they have been frequently misapplied.
>We need to really look at what robots should be used for and then work
>on the problems that specifically prevent them from being used in
>those applications.

And who will decide what robots "should" be used for?  Just because a
project doesn't seem to have immediate application does not mean that it 
hasn't advanced the state of the art and will not eventually find application
somewhere.  If we only concentrate on immediately-apparent applications then
we will fall behind in broader research, and ten years from now the Japanese
will be laughing at us (again).  This is not to propose that we ignore
applications.  Although I do think robots are most useful outside of 
manufacturing, many of the ideas that have been tossed around in conferences on
robotics could be applied (and have been, but in Japan or Europe) if the US 
manufacturing industry was more concerned with long-range planning than
short-term profits.

Nathan Ulrich                           "If it was easy, someone would have
ulrich@grip.cis.upenn.edu                done it already..."
DoD #0080 - GT750 pilot

jpexg@rice-chex.ai.mit.edu (John Purbrick) (09/06/90)

In article <5832@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) writes:
>.....I have never seen force 
>compliance used in an actual application, not because it is unavailable,
>but because it is not as needed as other technologies.
>
>......
>Mike "my favorite project was a robotic bartender" Smith
>HP Labs

Well, since you lead up to it--I made a system a few years ago which would 
visit a row of paper cups one at a time and pinch them with a force-sensing
servo controlled gripper. It could reliably tell the difference between single
cups and double cups (one inside another). The idea was to build a pyramid out
of the cups but I never got around to it.

Think how this could revolutionize the bartending field!

--John Purbrick

josip@ra.src.umd.edu (Josip Loncaric) (09/06/90)

In article <5832@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) writes:
>In article <29067@netnews.upenn.edu> ulrich@grip.cis.upenn.edu (Nathan Ulrich) writes:
>>
>>But I'm not a pessimist.  I think high-accuracy operations can be realized
>>with current technology and with serial robot manipulators.  But not with
>>precise position control and not with "endpoint" control, but with force
>>control.  This same low-accuracy human manipulation system can locate its
>>two hands relative to each other in space with enough accuracy to put a
>>0.9995" peg in a 1.0000" hole.  How?  By using compliance and force control.
>>This is better than any robot manipulator can accomplish without huge chamfers
>>and the use of a RCC (which is passive compliance).
>>
>>My opinion only.  Now tell me why I'm wrong.
>>
>
>I work for a $10 billion a year company and I have never seen
>an application where such a peg is put into a hole!  Such an application
>is about as realistic as bin picking or stacking legos.  It doesn't matter
>how well a system does something that is not useful.  This is a major
>contention of mine.  I believe that a great deal of robotics research
>is misdirected.  Hey, sure such stuff is fun.  My first paper was on
>bin picking and I have given demonstrations on compliance using a force
>sensor because it was  technically challenging and interesting but
>subsequently I found that it was misguided.  I have never seen force 
>compliance used in an actual application, not because it is unavailable,
>but because it is not as needed as other technologies.
>
>We have been collectively doing research in robotics for many years now
>and they are still not widely used in manufacturing.  Why?  Because they
>are not cost effective and because they have been frequently misapplied.
>We need to really look at what robots should be used for and then work
>on the problems that specifically prevent them from being used in
>those applications.
>
>Mike "my favorite project was a robotic bartender" Smith
>HP Labs

Well, perhaps you never had to put that size peg into that size hole...
but if you never had to deal with closely fit parts, I'd be very surprised.
The key point I'd like to make is that in assembling parts with very tight
tolerances (finer than the positional accuracy of the robot) you CAN and
SHOULD use force-based strategies.  Assembling parts by feel is much easier
than assembling them without force feedback.  This is another example where
designing parts so that their shape guides the assembly process through
force feedback can be helpful...  Anyway, the reason people have not used
active force control much   have to do with the technical difficulties of
attaching this gadget to a commercial robot - it's very hard to get
high enough bandwidth - but passive RCC devices are widely used.

Your second point (that robotics research is not being directly applied
to manufacturing) is well taken.  A classic example (from an MIT robotics
course) has students designing an elaborate two-handed dishwashing robot,
while dishwashing robots can be bought at Sears for $300.  So, what 
manufacturing needs are better manufacturing machines, and not "robots" in
the Capek's R.U.R. sense.  Many researchers, however, are motivated by
the problem of designing machines capable of functioning in unstructured
environments.  This has little importance in a highly structured production
process, but will be of great significance 10-20 years from now, as robot
applications become truly widespread.  These future robots may also help
reduce the complexity of learning production tasks in a factory (less precise
programming will be needed).

Summary: robotics research serves more goals than just manufacturing.
Force feedback (pasive and/or active) is a valuable technique in assembling
close fit parts, which should be designed to help the process along.

A final comment: HP probably uses robots mainly to stuff PCBs, which may
explain why you've never dealt with peg-in-hole type problems.  But, if you
assemble VCRs, Walkmans, servomotors, etc., this type of problem quickly
shows up.

--
Josip Loncaric / SRC / U. of Maryland / <josip@ra.src.umd.edu>
--------------------------------------------------------------
!            Today's Special: Opinions....$0.02 each         !
--------------------------------------------------------------

abg@stc06.ornl.gov (BANGS A L) (09/07/90)

In article <5832@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) writes:
>Mike "my favorite project was a robotic bartender" Smith

Well, just thought you might want to know that it has been done.  When I
worked for Honeybee Robotics, my primary project was the development of
just such a beast, using a GMF robot and various bits of automated
dispensing equipment.  We got some TV coverage, but so far, no nibbles
that want to spend money.  A couple of talk shows wanted to have the
robot on, but it was going to be a real pain to move it to a studio.

Alex L. Bangs ---> bangsal@ornl.gov         Of course, my opinions are
Oak Ridge National Laboratory/CESAR            my own darned business...
Autonomous Robotic Systems Group

ssridhar@pase60.Convergent.Com (Srinivasan Sridhar) (09/07/90)

During my research student days, we had several robotics projects involving
manufacturing and non-manufacturing tasks.  Precision was important in all
applications. Assembly line robotics required precise positioning of the 
manipulator.  Even sensing required precision positioning (as a result of
the sensing).  When we conducted industry sponsored projects, the specifications
put froth by most industries required accuracy (even the tasks demanded this).

For instance, using a manipulator for card assembly, force sensing
and positioning would require the system to be capable of micro-movements.

Surgical applications would most definitely require extreme positioning.

The applications go on and on and on............

___
sridhar

mgsmith@hplabsb.HP.COM (Michael Smith) (09/07/90)

In article <1990Sep6.202839.15676@cs.utk.edu> bangsal@ornl.gov (BANGS A L) writes:
>In article <5832@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) writes:
>>Mike "my favorite project was a robotic bartender" Smith
>
>Well, just thought you might want to know that it has been done.  When I
>worked for Honeybee Robotics, my primary project was the development of
>just such a beast, using a GMF robot and various bits of automated
>dispensing equipment.  We got some TV coverage, but so far, no nibbles
>that want to spend money.  A couple of talk shows wanted to have the
>robot on, but it was going to be a real pain to move it to a studio.
>

We used a robot to sell beer at 100% markup at engineering society events 
and then we would drink the profits.  It was fun.  Also got a bit of TV 
coverage.  Develop a system that revolutionizes manufacturing though and 
you are still a pariah.  

We have to get away from what people can do and have robots do  what we cannot 
(such as high accuracy assembly).  Such projects not being considered
as interesting is something we need to overcome.

Mike Smith
HP Labs

hbg6@citek.mcdphx.mot.com (09/08/90)

In article <5837@hplabsb.HP.COM> mgsmith@hplabsb.UUCP (Michael Smith) writes:
>
>We have to get away from what people can do and have robots do  what we cannot 
>(such as high accuracy assembly).  Such projects not being considered
>as interesting is something we need to overcome.
>
>Mike Smith
>HP Labs

Not considered interesting by whom? The popular press is still waiting
for the 100% accurate robotic lawn mower / window washer. A robotic
system which places a spindle in an interference fit hole just dosen't
make points in the ratings. If your playing to that audience, don't
hold your breath for interest.

John

hugh@ria.ccs.uwo.ca (Mr. Hugh Jack) (09/10/90)

I have only briefly noticed the on-going arguments in this
column about high accuracy manipulators, and it has 
reminded me about some discussion I have seen in papers.
The cornerstone of the problem can be described as
'Human/Super Human'.  The problem typically involves trade
offs between high accuracy and high flexibility.  As
humans we tend to concentrate on accuracy, flexibility,
power, or some combination thereof.

Many current robotics researcher have attempted to get the
best of both.  This is not to say that it is not possible,
but the best robotics systems we have seen to date
(i.e. humans) are never super.

Thus, I would like to bravely conclude that the question
of which is better or more useful is not the issue.  What
researchers should be moving towards is a system that is
able to use compromises between the different domains.

Hugh Jack
Graduate Student
University of Western Ontario
London, Ontario, Canada
hugh@engrg.uwo.ca

p.s. Anybody who takes these opinions seriously is as twisted as i am.