Arm model; scholarship

[From Bill Powers (920527.1000)]

Uwe Schnepf (920527) --

Welcome to the speaking population on the net, Uwe!

I am very much interested in the current discussion on arm control.

Good. The schedule for release of the model, roughly, goes like this. Greg
Williams is evaluating it now, and preparing some preliminary material for
an article on it, which Greg and I will submit to Science (and if they
don't take it, Behavioral Cybernetics). The article should be submitted by
the early part of July. I will be busy working up figures and adding a few
little refinements to the model, as well as working on the writing, until
then.

After the article is submitted, I will release the Arm Demo Version 2.0 or
2.1 for general consumption. The release will include an abbreviated
writeup without figures, summarized instructions, and complete C- language
source code with all special header files and object files needed to
compile it under Borland Turbo C version 2.0 or later. The target platform
is an IBM compatible 286, 386, or 486 machine using Hercules (720 x 348),
EGA (640 x 350) or VGA (640 x 480) graphics. I'd love to do a version for
other machines, but don't have the hardware (or expertise). The shareware
asking price will be $100 US. This crass commercialism is due to the fact
that I have no institutional or grant support for my work: everything comes
out of my pocket, which is currently empty.

I think you may find the model to be useful in your work, although it is
designed as an exploration of how the real human system is organized and
not just as a way to get a particular job done. As you no doubt realize, it
uses a hierarchical structure rather than a subsumption structure. It uses
binocular vision, which is easy to do in a simulation but much harder to
implement in hardware. Its motor actions are created by torques applied at
three frictionless joints. Its visual inputs and joints are positioned in
the anatomically realistic way (I'm not sure how the system would work with
the eyes mounted on the wrist!). Also, in my model there is no "triggering"
of any motor behaviors; all actions are smooth and the variables involved
in their control are continuous. Of course I realize that if grasping is
involved, there has to be timing of the initiation of various actions that
are appropriate only after certain control actions are complete.

A general question, to you and any other experts in robotics who are
listening in.

I need a way of simulating the movement of jointed masses that could be
extended to a model of the complete body. Intuition tells me that there
must be a simpler way to do this than by solving LaGrangian expressions
involving potential and kinetic energy. Doing it this way with a body
having four multiply-jointed appendages, a bending and twisting trunk, and
two swivels at the neck is far beyond my mathematical abilities.

In a simulation, it's easy to apply forces to a single mass and integrate
the resulting angular and lineal accelerations to produce
velocities and positions. The mathematics gets complex but the simulation
doesn't have to use analytical expressions: it just has to make the body
move realistically, for a short time, under applied forces. The
integrations don't have to be very accurate when control is involved; any
little errors are simply equivalent to small control errors, which are
self-correcting; in effect, initial conditions are re- established every
iteration.

I'm stuck, however, when it comes to doing this with even two masses joined
by a hinge or ball joint. I can't figure out how to express the effect of
the joint on constraining the motions of the masses. I know that there are
analytical ways to do this (I'm using one in the arm model), but that's
what I'm trying to avoid. I'm really trying to set up an analog-computer
statement of the problem, so the simulation itself can solve the equations
by "acting them out." I'm sure someone has done this. I'm not getting very
far trying to do it by myself. HELP!

If we had a way of setting up such a simulation, this would quickly provide
a basic structure to which control systems could be added, for a simulation
of far more than just one simple moving arm. The basic organization of Arm
v2. will work just as well for legs and a neck as for arms. The control-
system part is relatively easy to implement, at least approximately. The
hardest part is simulating the physics of the environment -- and as far as
the control systems are concered, the environment includes the body, which
is just a way of converting outputs into sensory signals.

I think we're close to being able to do realistic simulations of whole-
organism behavior. This one little (!) problem is all that stands in the
way. Anybody?
--------------------------------------------------------------------- Joel
Judd has claimed the remaining student scholarship for the meeting. If he
has to relinquish it (by the first of July, please) we'll immediately
notify everyone so someone else can have it. Couldn't think of a better
recipient.

ยทยทยท

--------------------------------------------------------------------

Best to all,

Bill P.