Machina Speculatrix and Emergent Behaviour

Grey Walter was ahead of his time. However, what you describe has
already been happening for a long time now, especially with the
subsumption-type approach to robotics developed by Rodney Brooks.
However, these sorts of bots cannot do much which is really
intelligent. Like Grey Walter's bots, they are basically limited to
operating in very specific environmental niches. They cannot deal with
the sort of general everyday problems that a 5 YO can deal with. They
cannot deal with abstractions, language, analysis of general visual
scenes, on and on.

Neither. Nor is robotics strictly a matter of reproducing ANY type of
behavior, though some platforms are built on the concept. Brooks admits
behaviors are perceived; they are abstract concepts that become whatever
we want them to be.

Grey Walter experimented with machines that had only a few functions,
because they were extremely simple mechanically. They could not right
themselves if tipped over. They could not coordinate travel over uneven
surfaces. They could not fetch a beer sitting right in front of them,
let alone from a refrigerator in a crowded house.

Behaviors are NOT merely what's programmed into the robot. This is a
common misconception. Behaviors are derived first from the physical
capabilities of the machine, just as they are in nature. Forget the
number of transistors in a microprocessor; that's a meaningless number
when robots are primarily mechanical devices. If a robot can't
physically climb stairs it makes no sense to build in behaviors for
this. Robots remain limited behaviorially because their mechanisms are
constrained by the laws of physics and $$$.

-- Gordon

 <snipped the good stuff>

    I have to agree.  My favorite robot architectures always include a
mixture of "reflexes" and programmed micros.  This method exhibits the
strong points of each type of construction.  For example, imagine a circuit
that consists of little more than a pendulum or microswitch that can detect
the robot being upside down.  Add to that a pair of solenoids or other
simple devices to set the robot upright, and send just enough data to the
brain to let it know that it had in fact been flipped over and when it is
back on its feet again, so to speak.
    Now, the flipped-over sensor constitutes a type of reflex, and the brain
only has to take into account the fact that the robot has been turned over,
and perhaps righted again.  You barely have to program that, and the
hardware will take care of itself.  You can also add other simple reflex
circuits that will report to the brain what has happened and do something
about it, and the brain simply does the bookkeeping then.  You add a couple
of lines of code or a loop or two for handling it, and it turns out that you
can very easily compartmentalize the code for that with little pain.
    I built a fire extinguisher bot a few years ago for the Trinity College
fire fighter competition that used a similar set of reflexes.  One used a
pair of fire sensor eyes aimed together at a point a few inches in front of
the robot.  When one eye saw the fire, it was used to navigate toward it.
When both eyes saw the fire, due to their angle and spacing, there was only
one place it could be- dead ahead at a fixed distance (it was a candle).  In
that unique case, the robot reflexively sprayed it with water.  The brain
only used the sensory data to steer by, but the eye/water squirter
combination would spray the instant it had the flame in sight.
    The brain could then do the "paperwork"- hey, we saw fire, headed for it
and since both eyes then got it, we successfully located it through
navigation.  Then, the reflex kicked in and told us so, and now there is no
fire- we must have put it out.
    So with a handful of simple sensors and reflexes, and a very minimum of
complex code, you get something that acts like an organism and exhibits
something similar to layered brain architecture.  I strongly recommend that
both hobbyist and professional try this approach and see how well even
thorny problems can respond to it.

Cheers!

Sir Charles W. Shults III, K. B. B.
Xenotech Research
321-206-1840

Since I didn't notice either Joseph Jones or Daniel Roth pipe in on
this thread, I thought I would  for them.  I gave myself  their book
"Robot Programming, a Practical Guide to Behavior-Based Robotics" for
Christmas.  I'm about halfway through it now.  It is a very good
introduction to the kind of systems Gordon and Charles have been
talking about.

Paul Pawelski

Sir Charles W. Shults III wrote:

....<snipped>

"Pierian Spring"

I've worked a little bit with swarm algorithms and even have a paper
published on ant colony simulation, and I do believe that emergent behavior
has more chances to succeed in tasks that need broad knowledge instead of
highly specialized knowledge.

It is easy to prove that a big number of simple agents (simple behavior)
interacting in great volume produce complex behavior, but it is very
difficult to develop a system based on emergent behavior that actually
produces a specific desired behavior. Complexity is difficult to control.

Cheers

Padu