Hod Lipson
<p>Jan 19, 2010 2:01 PM
 
Foresight Talk  (I chat with him)...  no repres language;  all done in neural nets
? cannot be modified or deciphered...  no bayesian stuff.
 
Overall impression  ??? is this real and generalizable  what is consensus abt this...
 
His area of interest and mine have huge overlap ...  need to read his papers ...
what is real and what is hype.  what is special purpose and what is generalizable.
 
He's interested in self-reflection and adaptive behavior;
unstructure environs;  crawlers  (see TED);  Rodeo  competitive array of controllers;
then simulate evolution to see which controllers are best;  then download those
into an actual robot to test it and collect sense data then put that back into the
simulation and repear.
 
SPIDER/ CRAB ROBOT  has 8 motors and 2 tilt sensors...
 
Imagine being in a CAR and not being able to see out at all (blind).
You must discern what your SHAPE IS; what appendages you have;
how you move; what you can do...  exactly like a BABY
great INTUITION PUMP!!! 
 
Do morphological estimation ...  what do I look like... do experiments to discriminate.
SCIENCE 2006 article;
 
Bongard and Zykov  inflated self-model  (just like academics)
Can it model its own controller?
Theory of mind:  modeling other people;
 
Adaptive morphology
self-replication
printable machine        Fab@home   print batteries;
robot creeates a model of itself;
do damage dx on a bridge by installing sensors;
AUTOMATING SCIENCE:  Eureqa  (sim to AM/ Heurisko)
 
Shake the system:  do new experiemtns to check out the local minima;
 
 </p>