Searching \ for '[OT] a robotengedanken experiment' in subject line. ()
Make payments with PayPal - it's fast, free and secure! Help us get a faster server
FAQ page: www.piclist.com/techref/index.htm?key=robotengedanken
Search entire site for: 'a robotengedanken experiment'.

Exact match. Not showing close matches.
PICList Thread
'[OT] a robotengedanken experiment'
1999\08\26@210536 by Anne Ogborn

flavicon
face
I propose, as a thought experiment, a robot contest.

 A series of typical small mobile robots are constructed
to a standard design - perhaps wheeled "turtles".

 One of these robots is equipped with a remote control unit.
The others are distributed to the contestants. The contestants
equip their robots with an autonomous control system and a set of
sensors. Contestants are forbidden to modify the mobility mechanism
(no fair making the robot fly, for example).

 A bare room is constructed with a solid floor and walls, except
for a "ditch", say 4' wide and 4' deep, which bisects the room
and which is not navigable by the robots.
Robots start on one side of the ditch, and navigate towards a
well_marked goal on the other.
The ditch is spanned by two replacable tiles, creating a binary
choice - to reach the goal by crossing tile A, or by crossing tile B.
Each tile covers half the ditch's length, so there is no question of missing
the tile entirely.

An assortment of these tiles, of varying construction, are available.

 A tile is placed across the ditch, the remote control robot is then placed in
the room and
driven to the goal by a human driver. Those tiles which the robot
crosses are called "crossable" and the others "uncrossable".

The two tiles are selected at random from an assortment, always one from
the crossable tiles and one from the uncrossable.

Each robot is repeatedly released into the room to cross the ditch, with
different sets of tiles.  The one which successfully reaches the goal the
largest percentage of times wins.

Now, my question is, how shall we design a robot to win this contest?
We have no prior knowledge of the nature of the tiles.
One might be bisected by a wall, another simply an open frame that
will drop the robot into the ditch, another might have a hole only
in one portion.

It seems that many robot navigation problems (and I freely admit my
robot building experience is limited) revolve around winning my little
contest.

And clearly there are some problems for which there is no solution.
Suppose you are designated to drive the remote control robot, and you
are confronted with two tiles which appear identical, both "boxes" that
form a barrier across their whole width, but with an entry and exit door.
What isn't known is that the floor inside one box is solid, the other
a trapdoor.

So, I invite you all, how does one construct such a robot?

--
Anniepoo
Need loco motors?
http://www.idiom.com/~anniepoo/depot/motors.html

1999\08\27@011628 by Russell McMahon

picon face
I can't see the point, yet, and I'm happy if there isn't one but -

OPTION 1

Use the right hand tile always - 50% success rate for random tile
selection. or

OPTION 2

Use the left hand tile always - 50% success rate for random tile selection.
or

OPTION 3

Generate random choice on each attempt - 50% sucess rate and fools tile
placer who has spotted the subtle algorithms used in versions 1 & 2.

OPTION 4

Now it gets harder :-)
Depends on how many trials you are allowed.
If number of trials is staistically small and tiles are cunningly designed
to fool an only moderately smart robot then option 1 to 3 may well be
better (Use the Force, Annie).
eg 2 tiles, one with a clear optical path but uncrossable, the other with a
zigzag path so no line of sight but crossable, would lead a "I can see
across  - it must be OK"decision maker to die every time.

Assuming the uncrossable tile does not give itself away and allow you to
leave after entry - ie you must commit to one or other and then enter, then
one could design various sensors to sense probable problems. IF the tile
designer is TRYING to trap you then he will win as he (she probably in this
case ;-)) has a brain and the robot is only programmed using a brain.

THUNKS

I reckon option 3 is liable to be pretty good :-)

I may have missed what you were driving at. Maybe a bit more description
would be useful.
Various assumptions need also to be laid out eg
- Can you only test before entry to tile.
- May you "save" yourself by leving an uncrossable tile if you detect it.
- Can extra sensors be mounted?
etc.

regards

               Russell McMahon




From: Anne Ogborn <spam_OUTanniepooTakeThisOuTspamNETMAGIC.NET>
To: .....PICLISTKILLspamspam@spam@MITVMA.MIT.EDU <PICLISTspamKILLspamMITVMA.MIT.EDU>
Date: Friday, 27 August 1999 13:06
Subject: [OT] a robotengedanken experiment


>I propose, as a thought experiment, a robot contest.

1999\08\27@023925 by Anne Ogborn

flavicon
face
> Now it gets harder :-)
> Depends on how many trials you are allowed.
> If number of trials is staistically small and tiles are cunningly designed
> to fool an only moderately smart robot then option 1 to 3 may well be
> better (Use the Force, Annie).

No, I think I implied that the # of trials is statistically large. And presumabl
y
the right/left position of the crossable tile is random.

> eg 2 tiles, one with a clear optical path but uncrossable, the other with a
> zigzag path so no line of sight but crossable, would lead a "I can see
> across  - it must be OK"decision maker to die every time.

>
> Assuming the uncrossable tile does not give itself away and allow you to
> leave after entry - ie you must commit to one or other and then enter, then
> one could design various sensors to sense probable problems.

no reason to think it would be so designed.  I was thinking of tiles with
more likely 'real world' situations - like, for example, a tile with a number
of office chairs sitting on it, or a tile with a number of those wire mesh
trash cans, or one littered with clothes.

> IF the tile
> designer is TRYING to trap you then he will win as he (she probably in this
> case ;-)) has a brain and the robot is only programmed using a brain.

that's not the objective, and of course making an uncrossable tile is as
simple as making the tile be "nothing" - the robot physically can't cross.
I'm not seriously suggesting we do this as a contest,
I'm suggesting thinking about it might lead to understanding some piece of
robot navigation.

{Quote hidden}

no. You can do anything physically possible except change the robot's locomotion
or add a human driver. Indeed, I'd be very suprised if the robot's algorithm was
n't
something along the lines of wall following. examine patches in front of you. If
you
think they're not safely passable, consider them walls and use a wall following
algorithm.

> - May you "save" yourself by leving an uncrossable tile if you detect it.

sure.

> - Can extra sensors be mounted?

yes.  I said that in the original post.

The physical setup seems pretty straightforward.
Here's a little robot, it needs to navigate from A to B, and in between is
known to be a path.  That's the navigation problem.
What I've done is reduce it to the problem of not getting stuck. Since there's n
o
requirement for speed getting to the goal, even an algorithm of moving randomly
into
'safe' areas will work.

1999\08\27@095252 by Dan Creagan

flavicon
face
OK.. I'll play a little bit (Mom said I could) 8).

I think most navigation problems are broken down to their various parts and
then conquered one at a time.  In this case, Annie describes what is
essentially an infinitely variable environment - if I interpret the rules
correctly.

If I'm right then just about anything could be in front of the robot - and
there is always a 50% chance of finding a successful bridge but not assured
discovery of a successful path on that bridge.  Could we come up with a
basic set of hierarchical responses to the environment and then model the
robot behavior accordingly?

Say:

1. preserve life (robot life)
2. explore
3. record and leave a record for others to follow (if I understood the rules
correctly).
4. cross the bridge (successfully)
5. procreate with other robots
6. become sentient
7. make the humans cross the damn bridge


Ok... ignore the last three.

Dan

{Original Message removed}

1999\08\27@110554 by Adam Davis
flavicon
face
Well, for a 100% success rate, the robot must have sensor to determine whether
it's current direction is going to cause trouble, sensors to detect the extent
of it's environment, and it must have a memory to determine where it has failed
before.

ie, assumeing the robot can be place anywhere on the one side, drive straight
toward the target.  During this time, scan the area around you to get a general
idea of where you are in the box, and whether you are going to go over the right
or left portion of the tiles.  If you run into an obstacle, back up to a
predetermined point (perhaps where you started) since you know which tile you
went over, you can aim for the other tile.

If one could place a ultrasonic sensor on a rotating turret on the robot, one
could determine the position it is in inside the box.  It would also be able to
tell the robot of obstructions in its path.  Using wire senors around the front
of the robot, pointing downward, it could detect a drop off.

The trapdoor is could be detected with sonar pointing down.  Cracks in the
tile(which would indicate a trapdoor) are 'visible' in sonar signals.
Unfortunately, the robot could succeed, where the human might not notice such a
setup.

I'm a firm believer that there is a sensor for everything...  The real question
is is how to use them effectively and efficiently.  The more you add, the more
power they require... etc....

-Adam

Anne Ogborn wrote:
{Quote hidden}

n the room and
{Quote hidden}

1999\08\27@132250 by Wagner Lipnharski

picon face
If you think that our body is a long experience adaptation to the
environment, you can think it is some kind of a highly developed robot.
This organic robot is pretty nice to supply examples when building
electromechanical robots. A great percentage of humans just don't go
where they can't see or understand (as a protection system), so a robot
could do something similar, if the sensors can't determine what is
ahead, just avoid it, or apply a sticker in the robot's front head
"Adventurer" and reprogram the unit to find out what is in the unknown
and unreadable by the sensors space ahead.

A memory (E2prom) mapped is the best choice, one byte per each 4 sq
inches (aprox 10x10cm) can save 256 different meanings for that space,
along with a nice electronic compass and space positioning by beacon
triangulation.  We use something similar based on space/time.  When
traveling you can have an idea how far you are and probably "where" you,
also where is North or South based on the movement/distance/time, if you
sleep during the travel you got confused and you can't guess where are
you (you can miss it by hundreds or thousands of miles, and you will
need a compass or other trick to say where is South).

A beacon can be used somehow with a transponder that works as a
retransmitting system.  Your robot send a signal (ultrasound or
something) and the fixed stations (three) will receive it and send
another signal back. Your robot will receive it and analyze distances,
positioning, etc.  Each station has a different preset delay to send
back the signal, so one will not interfere with another, they can use
the same signal frequency (if ultrasound). In real you would be
recreating a GPS system.  This is cheap and works, and your robot will
know exactly where it is in the mapped area. You need to consider
problems with signal bouncing, walls, objects, and so on, that can
interfere in the "understanding" of time/space. Measurements of arriving
signal amplitude and delays can tell the robot several other important
informations.

Pre-programmed objects and locations would be nice, as power outlets
(battery recharging), hiding and resting areas and so on.  The
understanding of the environment is pretty nice, as for example stay in
a corner during the night (connected to a power outlet) and working as
an automatic light upon noise (it can light your way to the bathroom
night run), or working as a "learning repeater unit", turning on the TV
and select a specific channel at certain time of the day, because it
"saw" it happens several times at its "infrared receiver", or "yell" to
the kids to go sleep because it is after 10pm.

Just don't forget to include the organic robots common mistakes as
"marriage", "get fat", and other minor specialties. :)

1999\08\27@201148 by Anne Ogborn

flavicon
face
Dan Creagan wrote:
>
> OK.. I'll play a little bit (Mom said I could) 8).
>
> I think most navigation problems are broken down to their various parts and
> then conquered one at a time.  In this case, Annie describes what is
> essentially an infinitely variable environment - if I interpret the rules
> correctly.
>

Dan - I think you've gotten to the root of it.
I don't see an answer to navigating an infinitely variable environment,
which seems to mean that there's not going to be a general rule for robot
navigation.

I know that's hardly staggering, but it's always nice to show there's no more
general case, even for the environments that theoretically CAN be navigated.


--
Anniepoo
Need loco motors?
http://www.idiom.com/~anniepoo/depot/motors.html

1999\08\30@165601 by John Griessen

flavicon
face
Hey,

Now you're thinking cleanly along the lines of the thought experiment
proposed
by Anne.  Memory of past success is one key to reducing the chance of
getting stuck
way down.  If you scan through past experience and see an environment that
fits the "impassable" rules of your capabilities, then go right to the other
tile/path/trial.

Has anyone read Society of Mind by Marvin Minsky?  His theory of memory is
via agents
that have inputs from other agents that have inputs form other agents and
connect up
all in parallel when one big mother agent is happy such that they can be
triggered again whenever the brain thinks of one of those happy outcomes.
"Thinks of"
means in light of short term memory around that past event, and
being one of the saved happy category of memories.  happy memories don't
collapse into
a single state.  They save their uniqueness in time and circumstance by the
thread of connection that can reactivate all the agents that participated at
the time of the event that was
remembered in context of causing a happy outcome.  To see if now matches a
happy or
"worn out trying" outcome, all those states are scanned through long enough
to compare to the states of "now" (including short term memory sequence
history).

Short term memory is key to any success in this kind of trial.

How do you make a set of short term memory agents to characterize
wall/obstacle edge
following?

Darn if I know...

John Griessen

{Quote hidden}

More... (looser matching)
- Last day of these posts
- In 1999 , 2000 only
- Today
- New search...