DPRG List  

[DPRG] New research maybe has some ideas for us ->Re: Hello again DPRG. (Re-Introduction) and long message.

Subject: [DPRG] New research maybe has some ideas for us ->Re: Hello again DPRG. (Re-Introduction) and long message.
From: Brad garton bgart at iadfw.net
Date: Thu Apr 19 11:01:55 CDT 2012

Maybe this  points to a part of the puzzle of making smarter mobile robots.




It suggest the brain places a penalty on long connections (similar on our
side I think to a ROS compute server somewhere else) but definitely uses
those connections, when needed. 


The scientists, from the Behavioral and Clinical Neuroscience Institute
<http://research.psychol.cam.ac.uk/~bcni/>  in the University of Cambridge
Department of Psychiatry, and the National Institute of Mental Health in the
U.S., discovered that the network can be modeled as a result of just two
different competing factors: a distance penalty based on the cost of
maintaining long-range connections between various brain regions, and a
second term modeling the preference for links between regions sharing
similar input.

Professor Ed Bullmore
<http://www.neuroscience.cam.ac.uk/directory/profile.php?edbullmore> , lead
author on the paper, explains the dynamic between the parameters they
identified: "There is a huge amount of evidence that the wiring of brain
networks tends to minimize connection costs. Less costly, short-distance
connections are much more numerous than more costly, long-distance
connections. So our model realistically includes a distance penalty on
long-distance connections, which will tend to keep connection costs low.

"However, we found that cost control alone was not enough to reproduce a
wide range of network properties. To do that, we had to model an economical
trade-off between cost control and another term which favoured new, direct
connections between regions that shared similar input or were otherwise
already indirectly linked."

So maybe it makes sense to add link cost to our local code/ROS code mix, if
it's not already there, and have  similar message structure between local
nodes and remote nodes for similar functions.  I could see for example
NavigateTo being the traditional  local behavior and a remote behavior using
the same message, topic  and service structure, but maybe using different
sensors and different compute resources internally.  Maybe the local one
uses geometry only to determine current heading, but gets stuck in canyons
etc. Remote uses SLAM for instance. New information kicks off the local
behavior and remote behavior at the same time, and  the robot gets commands
from NavigateTo (local), until NavigateTo (remote) asserts, possibly
overriding NavigateTo (local)'s waypoints  with new waypoints,  if
NavigateTo solves the problem first, maybe that's similar in effect to
"favoring" local connections. That's probably how you would write it anyway
but it's kind of interesting the brain seems to do much the same thing. 


If we only knew more about the compute side of Biology than we do now.  I'm
a ROS newb so I may be missing something important.  Right now I am not
using an architecture that keeps state and does learning on any local code.
Learning and room geometry  would have to be remote for my robot, until I
figure that out. 


BTW if you have more videos of the ROS talks can you post them to the web
site? I really liked the first one!



blackstag wrote: 

> I believe ROS is what you are looking for. It has a large foundation in
making the basics of robotics a simple task.

Jose I Quinones wrot: 

>I think that at the end, the successful robot will not be made of a
collection of behavioral algorithms, but a core of algorithms that allow
him/her to gather information and process in a way in which human-like
learning can be achieved. At the moment, there might be computers getting
close to doing this, but they take off entire rooms.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.dprg.org/pipermail/dprglist/attachments/20120419/195ed853/attachment.html 

More information about the DPRG mailing list