IK by relaxation or solving with newton and Danavit-Hartenberg

    It's just homework.

    There are analytical solutions for systems which have a
unique set of joint positions for each end effector position.  That's
what was used on industrial robots with very limited CPU power, and
many industrial robots are still doing it that way.  If your
robot has such a geometry, that's the way to go.

    If you have redundant degrees of freedom, so that there are an
infinite number of possible solutions, some additional problems
appear.  But CPU power for relaxation or minimization usually isn't
one of them any more.

    One problem with minimization methods is that, for some geometries,
you can get caught in a local minimum.  Think of reaching around to
scratch your back, but from the wrong side.

    Another problem is that, in practice, you want a sequence of
solutions, and you want them to be close together.  A small move
of the desired end effector position should result in small changes
in desired joint angles.  Solutions which don't have that property
are undesirable for robot control, since they'll cause unwanted
large joint moves.

                    John Nagle

Currently this is more of a theoretical question, since my solution just has
to show IK in action, but I would like to be able to easily extend it to a
very complex shape. The analytical solution will therefore not be an option.
The solution should easily adapt to different joint configurations, and they
should be alowed to be rather complex.


I am faced with selecting between minimization and relaxation, which in a
sence boils down to the same.


I am wondering. Can this never happens for relaxation? intuitively I would
expect it to be just as volnourable to that as minimization. After all,
relaxation is generally a linear minimization. If you imagine two joints
with a link length constraint, oriented so that they form a 45 degree sloped
line, and in between the two you have an irregularly shaped object. The
constraint is initially violated because the joint length is stretched out,
so the joints need to move together. As they do they collide with the object
and get locked into an indentation, and relaxation wont get it loose..
unless you add some random wiggeling.


That makes sence. Will minimization and relaxation not both give the
smallest movement, if the starting point of the methods are both the initial
position?

Can you confirm my understanding that the initial text does not in any way
implie that relaxation can not be used?
Another thing... the relaxation I have done is not related to robotics, so
perhaps things are done diferently. The very basic idea is to have joints
connected with distance constraints. You start with an initial pose, you
move the parts thet you want moved and fix them. Then the constraints are
broken and you relax the system, which will pull the non-fixed parts in
closer so the constraints are not violated.
The reason I keep digging in that specific subject is that looking for
robots and relaxation never returns anything but into on route planning and
other non relatex subjects.

If your kinematic chain has some extra DOF you must select one of family of
solutions as sole output (result). In case if such selection is hidden (implicit)
part of compared approaches, you must expect different output from different
methods. If criterion of selection of single result from set of possible is
explicitly formulated and criterion is the same in both cases you will get
the same result.

I am not sure that I understand you. Are you saying that if there is only
one solution (which is rarely the case) thebn both methods will give the
same result, whereas a multisolution system will give different solutions,
so I need to decide which kind of solutions that i want?
Will relaxation and minimization (by newton with the jacobian) with the same
starting and ending pose, not always give the same result?

Yes. But maybe in your case any solution is appropriate?


Moreover, if minimization algorithm has some parameters then processes with
different values of such parameters will produce different solutions in case
multisolution.

In case if multisolution is presented A PRIORI then you can select particular
unique solution by adding extra term to minimized function which
expected to be minimal for most desired solution (for example sum of squares of
joint position increments or energy consumed for motion from previous state).
Multiplier for such additional criterion is an additional parameter of algorithm
and must be assigned after experimenting; when this factor has appropriate value
then result of minimization is insensitive for value of multiplier in relatively
wide range.

In multisolution case Jacobian has one or more zero eigen values so adding
correspondent eigen vector to particular solution with any multiplier produce
another (close) solution; additional criterion used to select most appropriate
solution from such family.