From Greg & Pat Williams (920525)
In a word, Bill, your Arm Demo 2.0 is IMPRESSIVE!
We put the program through its paces (or rather, pointings) yesterday (yes, it
rained!) and found that the arm trajectories after correction of kinesthetic
space are very much like empirical data -- much, much more so than in version
1 (visual endpoint control). For some start/end point combinations, the "in"
vs. "out" trajectories are still a bit farther apart than the data show, but
this is a minor difficulty apparently related to head and eye motion which can
be fixed later, no doubt. Overall, the model is eminently reasonable from the
points of view of development, perceptual adaptation, and studies of people
with proprioceptive neuropathies (Bill, tomorrow we're mailing you a paper on
the latter which you should find quite supportive -- G. Ghez, et al., "Roles
of Proprioceptive Input in the Programming of Arm Trajectories," COLD SPRING
HARBOR SYMPOSIUM ON QUANTITATIVE BIOLOGY 55, 1990, 837-847). It is truly
remarkable for a model to meet the data in detail AND make sense in a general
way. Congratulations!
That said, we are assuming that the documentation as it stands is for us beta
testers only, and that it will be improved for reviewers and users. We think
you should start off with a detailed explanation of the model (incorporating
material like that in some of your CSGnet posts of last month). We're
available as editors if you're game.
It would be nifty to be able to explain the idiosyncracies of particular
trajectories (i.e., why that one curves convexly, not concavely, but has a
little concavity at the far, not the near, end, etc.) via an examination of
the arm geometry constraints/corrected kinesthetic space. It would literally
blow Bizzi's mind, we suspect, if you could say (in advance, for any
reasonable set of endpoints) how a trajectory would be idiosyncratic, and why.
But that is for the future.
More integer arithmetic might be welcome to those without coprocessors. Yep,
it is a pain in the posterior, but it ran only 1/10th real time on Pat's 16MHz
386SX. (A lovely 1/3 real time on my lowly 12 MHz 286 -- but with a souped up
IIT 287!) Again, for the future. "Real scientists are patient."
Why don't the visual and mapped modes work well together? How do you plan to
make them cooperative (or at least coexistent)? We assume that this will
involve some fundamental postulates about whether vision just notes the end
points or "helps" continuously. There are reports of more experiments on this
besides the paper we are sending tomorrow.
If SCIENCE doesn't like this, BIOLOGICAL CYBERNETICS should!
Your labors have indeed been well-directed! We'd say "point on," but it would
be a lousy pun....
Greg & Pat