From fredm  Wed Apr 23 02:22:07 1997
Received: from aleve.media.mit.edu by hub.media.mit.edu; (5.65v3.2/1.1/06Jun95-8.2MPM)
	id AA15010; Wed, 23 Apr 1997 02:22:07 -0400
Received: from ml.media.mit.edu (ml.media.mit.edu [18.85.13.107]) by media.mit.edu (8.7.5/ML961206) with SMTP id BAA25988; Wed, 23 Apr 1997 01:02:38 -0400 (EDT)
Message-Id: <199704230502.BAA25988@media.mit.edu>
X-Authentication-Warning: aleve.media.mit.edu: Host ml.media.mit.edu [18.85.13.107] didn't use HELO protocol
To: dahlgren@trincoll.edu, jmendel141@aol.com
Cc: handyboard, bss, ocko, meerse@agate.net, zagrynt@hotmail.com
Date: Wed, 23 Apr 97 01:02:37 -0400
From: "Fred G. Martin" <fredm>
X-Mts: smtp

AN OPEN LETTER TO PROFESSORS JAKE MENDELSSOHN AND DAVID AHLGREN,
ESTEEMED FOUNDERS OF THE TRINITY COLLEGE FIRE-FIGHTING ROBOT COMPETITION.

						Wed Apr 23 00:50:42 1997

Dear Jake and Dave,

Thank you for organizing such a fantastic robotic event.  It was my
great pleasure to have been an invited speaker at the fourth annual
Trinity College Fire-Fighting Robot Contest.  The event is more than a
contest; it is a meeting place for robot enthusiasts of all ages.  I
particularly enjoyed meeting the elementary school-aged children, and
their parents, many of whom I previously knew by e-mail only.

Now allow me to play the role of unsolicited trouble-maker, by making
some observations about the contest itself, and the types of robots it
encourages by the structure of the contest rules.  As I discussed in
my presentation, I believe that robotics is most interesting when the
robot is thought of as a creature-like entity rather than a machine
following a predetermined sequence of actions.  My robotic entry in
the contest (co-designed with Steve Ocko and Brian Silverman) tried to
embody this philosophy of robotics: it had a collection of behaviors,
all operating simultaneously, that attempted to perform the task of
putting out the candle's flame through their interactions, some
planned and some accidental.  For the record, our design did indeed
accomplish this goal, succeeding in two of the three official contest
runs.

On the surface, what was distinctive about our robot was that it was
the absolute *worst* of the robots that were successful in putting out
the candle.  That is, because our robot wandered around the playing
field in an organic fashion, it took lots longer than the others to
find the flame.  Our robot also hit the walls repeatedly, accruing
lots of penalty points, but even without the penalties, it was
squarely in last place of the successful robots.

However, our robot demonstrated two qualities that the contest should
encourage in other designs.  First, our robot was the only machine
(that I know of) that was designed to recover from errors.  In each of
our three runs, things went wrong and our robot struck the wall and
appeared to be stuck, but in all cases, our machine got itself unstuck
and continued along its merry way.  This was highly unusual---when
most other robots struck the wall, they were done for.

The other characteristic feature of our design that should be
encouraged is that we did not program the maze into our robot's head;
rather, we gave it a set of behaviors that would allow it to negotiate
*any* maze. (Well, sort of: our strategy was left-wall-following,
which if executed perfectly, would only bring us through three of the
four rooms in the contest maze.)

On the other hand, many if not nearly all of the other robots were
programmed with specific information about the specific maze they
would face, typically as a series of path segments, rotations, and
candle-searches.  While this may be a stout approach for the task at
hand, it does not generalize to other situations.  The puzzle you gave
us came with a map, but in any imaginable real-world situation, the
robot could not possibly have a map for all environments it would
encounter, especially since places populated by people are constantly
changing.

I would like to mention one other unusual robot, an entry in the
junior division by Jessica Anderson (Jessica was one of my students).
Jessica's design was particularly simple.  It had but three sensors: 2
touch sensors and one light sensor (for seeing the candle's flame).
The robot itself was a two-motor "turtle" design that carried a fan
powered by third motor.

Jessica's robot basically wandered around the maze, backing up and
turning when it bumped into the wall. Occasionally, it would turn left
and right, looking for the candle, but it was able to react to the
candle at any time. If it saw the candle, it would stop moving and
turn on its fan.

In short, Jessica's robot relied on randomness in order to work. What
happened in the actual contest? In her first run, the robot made a
series of right hand turns, working its way around the detached
room---and finding its way out the starting door! Jessica quickly
picked the robot up and put it back on the starting position, hoping
it would be allowed to continue. This seemed to confuse the judges;
they hadn't seen this happen before and it took them a little while to
realize that Jessica had violated the "no touching your robot" rule.
So the first round was ended.

On the second round, Jessica's robot was accidentally started with a
test program that made it just drive straight ahead; after plowing
into a wall, that round was ended.  But on its third run, Jessica's
robot proved its mettle, finding the candle, turning on its fan, and
putting out the flame!  Jessica's design proved that you don't need a
plan to be successful; you simply need persistence and good luck.

(Incidentally, on the first of its three contest runs, our robot found
itself in the detached room that it wasn't supposed to be able to
find, through a pair of self-canceling "errors."  On that particular
run, the candle was indeed in the room, but our robot was not able to
see it because it only had light sensors facing the opposite
direction---the way it was "supposed" to be able to find the candle!
If we had done a better job designing our robot to take advantage of
serendipity, we might have had a perfect 3-for-3 record this year.)


With this all in mind, allow me to put you guys on the spot by
recommending some specific changes for next year's contest:

REMOVE THE PENALTY POINTS FOR STRIKING THE WALL. I think that part of
the reason why so few robots are able to recover from hitting the wall
is that the rules penalize a robot for doing so.  I know that I fell
into the trap of thinking "well, my robot is not designed to ever hit
the wall, so I won't worry about what happens if it does" in our
original entry last year.  This year, we specifically acknowledged to
ourselves that we would rather hit the wall, accept the penalty, and
move on than hit the wall and die.  Jessica's robot hit the wall as a
matter of its normal operation, and that didn't hinder its
effectiveness.

Furthermore, removing the wall penalty probably wouldn't change the
final rankings of the robots much. Because recovering from hitting the
wall takes time, our robot was still squarely in last place among the
successful robots based solely on the raw time score.

But removing the penalty for hitting the wall would encourage
robot-builders to make designs that incorporated touch sensors to
recover from such contact, as well as simple designs like Jessica's
that roam around using touch sensors, a simple but effective
technology.


RANDOMIZE THE ROBOT STARTING POSITION. If it were not for the
practical difficulty of re-configuring the maze, I would suggest
creating a new random maze for each robot run in the contest.  But
establishing several robot starting positions and choosing one at
random before each run accomplishes the same goal: encouraging robot
builders to create designs that can solve mazes-in-general, not just
a-specific-maze.  This is really critical to encourage builders to
think about maze-traversing algorithms or effective wandering
strategies rather than pre-programmed path sequences. This change
alone would have the greatest influence in encouraging robots that
truly took into account their surroundings, and possibly could have
the side-effect of yielding more reliable designs overall.


CLOSE OFF THE EXISTING START POSITION WALL. This is a minor point, but
to encourage robots that can wander around with out fear of falling
out of the playing field, it would be helpful to make the maze a
completely enclosed space.


KEEP THE SCALE OF THE MAZE THE SAME. Dave, I heard you say that you
were thinking about scaling up the playing field. Please don't do
this. The current scale---robots up to a cubic foot in size---is a
wonderful size for robot enthusiasts. For example, consider this
year's AAAI set of robot contests: many of them require robots to be
substantially larger, putting them out of the realm of accessible
construction materials like the LEGO system.


MAKE THE ROBOT SIZE RULE "AS MANY ROBOTS AS YOU CAN FIT INTO A SQUARE
FOOT," NOT JUST "ONE ROBOT THAT MAY BE NO LARGER THAN ONE SQUARE
FOOT." OK, this is a bit forward-looking, but why not encourage
multiple-robot solutions?  In designing the contests for the MIT LEGO
Robot Design Competition, there was a dilemma with regard to multiple
robot solutions: I wanted to encourage them, but we gave students only
one CPU per robot kit, and they were not allowed to use more.  This
meant that if a team were to create a multiple robot design, it would
have to have a wire bundle connecting the remote robot to the robot
that carried the brain.  Many dual-robots with wire bundles tangled
themselves and the other robot up with the wire bundle, and we were
forced to outlaw multiple-robot designs out of a sense of fairness to
the opponent that would face their wire bundle.

But there are no such restrictions or complications in the
Fire-Fighting Contest.  I have always seen wonderful ideas come up
when students consider how two or more robots might work in tandem to
accomplish a given task.  Especially with the popularity of tiny CPU's
like the Microchip PIC series, why not encourage Fire-Fighting Robot
builders to consider these possibilities?  Certainly we can all agree
that cooperating robots will be part of our future.  By the way, make
sure to not bifurcate the contest into single- and multiple-robot
categories; the challenges of getting more than one robot to cooperate
with its partners should offset any numerical advantage that a team
that fielded a multiple-robot entry might have.  And there still is
the same volume limitation.


At first glance, it might seem like the changes I am suggesting would
make the contest infinitely harder: if robot designers don't know
where their robot will start, they can't design their machine to
execute a specific series of motions.  But so many robots failed
because this dead-reckoning approach is so hard to pull off.
Paradoxically, I believe that by making the contest "harder," it will
make it easier for us robot-builders to conceive of solutions that
will be genuinely effective.

Jake: in your comments that opened the contest event, you suggested
that we are witnessing the equivalent of the Wright Brothers first
flight; primitive robots that are precursors of the genuine household
robots of the future.  Well, how about giving us a contest that pushes
us beyond the path-sequencers that were so prevalent this year!


OK, this is all for now!  Dave and Jake, I know that I am putting you
on the spot by publishing this letter.  I hope that you read this
message as my desire to prod a great thing to be even better.  I truly
appreciate the work that you two have done in building this contest
and community event---if you change nothing at all, I will be back
next year anyhow!  But I hope you consider the ideas I've suggested
here.

Yours sincerely,
Fred Martin
fredm@media.mit.edu