[From Bill Powers (931213.1010 MST)]
Tom Bourbon (931213.0910) --
A number of people to whom I have demonstrated Little Man were
more interested in the visual control systems than the
kinesthetic. On the other hand (no pun intended), to what
degree would you eliminate the visual control systems? Aren't
they essential to the demonstration -- the image of the target
must be kept on the "fovea," must it not?
To show the visual control systems, all you need is Little Man
Version 1, which assumes perfect lower-level control systems. In
addition, you get the live block diagram.
The problem with the visual systems in Little Man V. 2 is that
they have a different center from the kinesthetic systems -- the
coordinate system is eye-centered instead of shoulder-centered. I
haven't been able to get the gain high enough (with stability) to
compensate for the difference in coordinate systems for fast
movements (for slow movements there's no problem).
This is why I was playing around with the mapping concept. If a
step-error in visual position is converted to nearly the right
step-change in target position for kinesthetic control, the
unwanted movements under combined visual-kinesthetic control go
away. It's equivalent to suddenly presenting the kinesthetic
control systems with a new reference signal which then remains
steady. That is what produces movements that match human
movements with pretty good fidelity. The data we have concern
movements that take roughly 400, 800, and 1200 milliseconds from
start to stop; by adjusting the integration factor for
kinesthetic level two, all three conditions can be matched,
including the path curvature details (which is interesting in
itself). So at these speeds, it seems that the observed action is
done without much effect from visual feedback; it would be nice
to have some data for the same movements done in the dark.
What I would really like to do is to put the map on the
perceptual side instead of the output side. This would make the
visual perceptual space coincide with the kinesthetic space.
However, when I tried doing this, I didn't have a good way of
building the map. The changes in the map should be gradual and
spread out over space, so that during adaptation there would be
no sudden jumps from one cell of the map to adjacent cells. My
way led to all sorts of discontinuities, which screwed up
control, which screwed up the reorganization process. This whole
part of the model needs to be redesigned.
···
--------------------------------------------------------------
On another subject: At the Cleveland VA Hospital, a doctor named
Byron Marsolais has been working for several years to implant
muscle electrodes in people who have had spinal-cord damage, to
give them a way to walk. So far he has used only open-loop
control. I wrote to him once in 1991 after seeing a PBS report on
his progress, and again this fall after seeing an update
(Scientific Americal Frontiers). Both times I urged him to try
putting feedback sensors on the joints, but he didn't reply to
either letter. The results he is getting are grotesque, although
of course any kind of walking is a joy to the people who had been
paralyzed. Marsolais has to drive opposing muscles to the maximum
in order to get any kind of control, with the result that
movements are big and jerky, and the patients become exhausted. I
pointed out that by using joint-angle feedback, he could produce
smooth control of positions, with automatic adjustment for
varying loads and mechanical advantages, but so far to no
apparent avail.
So I was wondering if there is anyone at the U of Texas Medical
School trying to do anything similar, and if there might be any
interest in developing a control-system approach to the problem.
---------------------------------------------------------------
Best,
Bill P.