Loading ...
Sorry, an error occurred while loading the content.

Re: [beam] Re: Outer Space Robots WOOT!

Expand Messages
  • David Buckley
    Martin You read my mind! However tonight a friend showed me his new PIC board with a tiny 6 pin smtPIC also available as an 8 pin DIP. It has 2 PWM, 1 A/D,
    Message 1 of 23 , Oct 21, 2012
    • 0 Attachment
      You read my mind!
      However tonight a friend showed me his new PIC board with a tiny 6 pin smtPIC also available as an 8 pin DIP.
      It has 2 PWM, 1 A/D, software selectable clock - 31KHz - 32MHz, an 8x8 user configurable ASIC, has 25mA i/o capability and runs at 50microamps.
      At present he is writing a macro library for it. No I don't remember the type number!!!
      He showed it bidirectionaly controlling a motor (through a H bridge) and current sensing stall at each end of travel to reverse direction. With a split battery supply he could do two motors. The downside is you need a PicKit-3 to program it.
      Somewhere between it and the PIC32 it must be possible, taking inspiration from nature, to figure out connection schemes which will allow robots to do much more than stumble forwards and back, and do it in some understandable way which is expandable.
      ----- Original Message -----
      Sent: Friday, October 19, 2012 8:26 PM
      Subject: Re: [beam] Re: Outer Space Robots WOOT!


      True, and true.  But, the question ( or my question anyway ) is how we can most effectively organize the neurons.

      The approach that most connectionist researchers ( neural network models ) have taken is to use "learning rules" to organize the networks based upon the the data they are presented.  This works surprisingly well, with some large networks ( >100,000 nodes ) being able to compete with humans at handwriting recognition tasks ( and other, similar, pattern matching "stuff" ).  The networks are all but opaque to deciphering how they work though.  Beyond that, the learning procedure is fragile.  It is very easy to construct a data set that leads to networks that perform badly or fail to generalize ( the, so called, overfitting problem ).  

      One of the reasons for mapping the neural networks of simple organisms is to gain insight about how to create different systems with the types of neurons available.  That is simplified by finding groups that work together and making modular "cuts" there.  But, in the end, it is a question of complexity.  How much can we understand "at a glance?"

      Looking forward, I would think it would be pretty easy to implement simple enough neurons and memory elements that I could fit 100s of neurons and 1000s of connections into a low power microcontroller ( a 32-bit ARM ).  Honestly, the biggest constraint is likely to be the amount of SRAM available on the chip.  If many of these modules could be used on a robot, there could be 1000s of total neurons -- nothing like a house fly ( down by several orders of magnitude ), but, potentially, a huge advance from current BEAM networks.

      Martin Jay McKee

      On Fri, Oct 19, 2012 at 11:27 AM, David Buckley <david@...> wrote:

      So little neurons!!!
      As I said before an ordinary housefly has about 1 million.
      That's 100,000 times as many as a complicated BEAM robot. And BEAM robots and Processor based robots don't have the thousands of neuronal interconnections and hundreds (thousands) of sensors, so we have a long way to go.
      ----- Original Message -----
      Sent: Friday, October 19, 2012 5:31 PM
      Subject: Re: [beam] Re: Outer Space Robots WOOT!


      You sure onto something, tiny insects can do so much with so little neurons, its the way they are put together and how they learn including what is already in there, such as instinct for different creatures, hence lots of animals can walk immediately etc.


      From: G.A. Anderson <autisticando@...>
      To: beam@yahoogroups.com
      Sent: Friday, October 19, 2012 10:15 AM
      Subject: [beam] Re: Outer Space Robots WOOT!

      I'm new here, but I would just like to add my $.02.

      I have a M.A. in Experimental Psychology, and as such I know a bit about how the brain is. Not a tremendous amount, but I still know more than most.

      I don't think that it is the amount of neurons in gross that will determine the complexity and survivability of a BEAM robot, but rather the patterns between the neurons. This is good news all around because complex BEAM robots do not necessarily need to be a large as we once previously thought (or at least this is possible). Right now, the circuit components that make up the BEAM robot's "brain" are much larger than virtually any neuron, but this may change in due time. As such, we should focus on constructing the greatest amount of patterns with a relatively lesser number of components.

      Another thing is that a BEAM robot has one innate advantage to a program-controlled one: it is much more similar to something that is actually alive. Some of the greater BEAM robots seem to have minds of their own, with their behaviors somewhat unpredictable. With a program-controlled robot, all it could really hope to be is a simulation. Programs are really just that: simulations. I could go on more about it, but I'm running out of energy.

      Anyways I hope that I said some insightful things. I can barely even take motors out of CD players so obviously I lack experience, but I nonetheless think that I am onto something.


    Your message has been successfully submitted and would be delivered to recipients shortly.