Now Available
~~~~~~~~~~~~~
ARS MAGNA
The Abstract Robot Simulator
Sean P. Engelson
Department of Computer Science
Yale University
ABSTRACT:
~~~~~~~~
AI planning research has historically operated in formal abstractions
of the real world. This approach was useful in discovering many
fundamental issues underlying planning; also, problems in simple
domains such as the blocks world can turn out to be surprisingly
difficult. Lately attention has turned to planning for more realistic
domains in which micro-world simplifying assumptions do not hold.
This shift of focus introduces a new problem of validation and
comparison of different planning theories and systems. A proper
domain for planning problems must be realistically complex but also
simple enough to support controlled experimentation.
To address these questions, we developed the ARS MAGNA robot
simulator. The simulator provides an abstract world in which a
planner controls a mobile robot. Mobile robotics is a particularly
apposite domain since it is a major application area for AI planning
techniques. ARS MAGNA's environment and robot models are based on
current robotics research, so that the domain is reasonably realistic.
At the same time, we abstracted away from many (though not all)
real-world details of kinematics and motor control. Experiments may
be controlled by varying global world parameters, such as perceptual
noise, as well as building specific environments in order to exercise
particular planner features. The world is also extensible to allow
new experimental designs that were not thought of originally. The
simulator also includes a simple graphical user-interface which uses
the CLX interface to the X window system.
DOCUMENTATION:
~~~~~~~~~~~~~
Version 1.0 of the ARS MAGNA simulator is documented in Yale Technical Report
YALEU/DCS/RR #928, "ARS MAGNA: The Abstract Robot Simulator". This report is
available in the distribution as a Postscript(tm) file, as well as from
Paula Murano
Yale University
Department of Computer Science
P.O. Box 2158 Yale Station
New Haven, CT 06520-2158
Email: ······@cs.yale.edu
Any and all comments would be most welcome, and may be directed to me, at
········@cs.yale.edu.
AVAILABILITY:
~~~~~~~~~~~~
ARS MAGNA can be gotten by anonymous ftp from ftp.cs.yale.edu, as
ars-magna.tar.Z in the pub/nisp directory, as follows:
% ftp ftp.cs.yale.edu
Connected to dept-gw.cs.yale.edu.
220 ra FTP server (SunOS 4.1) ready.
Name (ftp.cs.yale.edu:engelson): anonymous
331 Guest login ok, send ident as password.
Password:
230 Guest login ok, access restrictions apply.
ftp> cd pub/nisp
250 CWD command successful.
ftp> bin
200 Type set to I.
ftp> get ars-magna.tar.Z
200 PORT command successful.
150 Binary data connection for ars-magna.tar.Z (128.36.17.10,1220) (528589 bytes
).
226 Binary Transfer complete.
local: ars-magna.tar.Z remote: ars-magna.tar.Z
528589 bytes received in 7.4 seconds (70 Kbytes/s)
ftp> quit
221 Goodbye.
% uncompress ars-magna.tar.Z
% tar xf ars-magna.tar
%
Installation instructions are in the file Installation.readme. The
simulator is written in Nisp, a macro-package for Common Lisp. Nisp
can be retrieved in the same way as the simulator.
--
Sean Philip (Shlomo) Engelson
Yale Department of Computer Science
Box 2158 Yale Station
New Haven, CT 06520
·············@cs.yale.edu (Sean Philip Engelson) writes:
> Now Available
> ARS MAGNA
> The Abstract Robot Simulator
>
>... Lately attention has turned to planning for more realistic
>domains in which micro-world simplifying assumptions do not hold.
>simple enough to support controlled experimentation.
>ARS MAGNA's environment and robot models are based on
>current robotics research, so that the domain is reasonably realistic.
>At the same time, we abstracted away from many (though not all)
>real-world details of kinematics and motor control.
Despite being called a "robot simulator", this is a world modeller
without much reality. It's a 2D grid world, and it is basically
predicate-oriented,
not geometry-oriented. The robot is modelled as a point with "sensors",
"effectors", and "internal carrying capacity", in a world with "walls"
and "things". You code a planning program in "nisp" (a LISP dialect)
which does such things as calling "hand-grasp-desig", commanding a
designated hand to grasp an object. This succeeds or fails, depending
on the contents of nearby cells, and sometimes fails randomly, so the
planner has to be able to recover from errors. That's the lowest level
of detail supported. No kinematics, no dynamics, no gravity, no inertia,
no friction, etc.
John Nagle
In article <·····················@netcom.com> ·····@netcom.com (John Nagle) writes:
>·············@cs.yale.edu (Sean Philip Engelson) writes:
>> Now Available
>> ARS MAGNA
>> The Abstract Robot Simulator
[Much deleted]
> Despite being called a "robot simulator", this is a world modeller
>without much reality. It's a 2D grid world, and it is basically
>predicate-oriented,
John Nagle's criticisms are valid. However, it should be noted that
the fact that actions can fail, that sensing is limited, and that failures
can occur randomly, puts this simulator head and shoulders above a great
many simulators which are regularly promulgated in the AI literature.
Furthermore, an accurate geometric simulation of a mobile robot in an
unengineered environment is EXCEEDINGLY difficult to produce, especially
if you want it to run in something approaching real time.
IMHO.
Erann Gat
···@robotics.jpl.nasa.gov
···@aig.jpl.nassa.gov
In article <·····················@netcom.com>, ·····@netcom.com (John Nagle) writes:
|> ·············@cs.yale.edu (Sean Philip Engelson) writes:
|> > Now Available
|> > ARS MAGNA
|> > The Abstract Robot Simulator
|> >
|> >... Lately attention has turned to planning for more realistic
|> >domains in which micro-world simplifying assumptions do not hold.
|> >...simple enough to support controlled experimentation.
|> >ARS MAGNA's environment and robot models are based on
|> >current robotics research, so that the domain is reasonably realistic.
|> >At the same time, we abstracted away from many (though not all)
|> >real-world details of kinematics and motor control.
|>
|> Despite being called a "robot simulator", this is a world modeller
|> without much reality. It's a 2D grid world, and it is basically
|> predicate-oriented,
|> not geometry-oriented. The robot is modelled as a point with "sensors",
|> "effectors", and "internal carrying capacity", in a world with "walls"
|> and "things". You code a planning program in "nisp" (a LISP dialect)
|> which does such things as calling "hand-grasp-desig", commanding a
|> designated hand to grasp an object. This succeeds or fails, depending
|> on the contents of nearby cells, and sometimes fails randomly, so the
|> planner has to be able to recover from errors. That's the lowest level
|> of detail supported. No kinematics, no dynamics, no gravity, no inertia,
|> no friction, etc.
We'd like respond to Nagle's comments point by point. A detailed
justification of the ARS MAGNA model would be too long here; a paper
is in preparation which does so, placing the simulator in the context
of other current research. It should be noted that ARS MAGNA is
intended to simulate a world, robot, *and low-level sensing/control
systems*. The models used are based on current work in vision, sonar,
and behavior-based control.
- Nagle's core criticism seems to be focused in the first sentence,
where he accuses ARS MAGNA of being "without much reality". However,
while his distinction between "robot simulator" and "world modeller"
may be a real one, there's clearly a continuum; where one draws the
line (indeed, whether or not doing so is even useful) is a matter of
opinion. Hence this comment appears to be little more than a semantic
quibble.
- ARS MAGNA is not a "grid world" in any, except a very weak, sense. In
Section 2.1 of the manual it is made clear that the robot does not move
or sense on a grid, these points being where a grid approximation does the
most violence to realism.
- As to the claim that the simulator is "basically predicate-oriented"
vs. "geometry-oriented" (the latter is to be preferred, we gather?),
we find this claim curious. ARS MAGNA is "basically" neither; there
are geometric aspects as well as predicate-like aspects to the system.
Please read Sections 3.2.2 and 5.4 as well as pp. 19-26 for many
examples of non-"predicate-oriented" behavior. Perhaps there is some
other precise meaning of "predicate-oriented" of which we're unaware?
- Nagle's description of the operation "hand-grasp-desig" betrays a
misunderstanding of a fundamental concept, that of designation. No
robot commands operate on "objects" (other than robot parts). Rather,
the concepts of visual marking/tracking and servo control are packaged
in the concept of "designators" (see Section 3.4.1 of the manual). A
designator is a perceptual marker containing sufficient information
for a control module to perform its function, say, to servo to some
location. The error models here are realistic, assuming implementation
of this reasonable low-level model.
We believe that the problem here mainly stems from a confusion about
what sorts of problems are being addressed. If one is working on
low-level control issues, ARS MAGNA is clearly inappropriate, as it
simulates at a higher level of abstraction. If one's domain of
interest is task-level planning or large-scale mapping, ARS MAGNA is
eminently appropriate, and, as Erann Gat mentioned, rather better than
other simulated worlds available at this time.
Sean Engelson
Drew McDermott
--
Sean Philip (Shlomo) Engelson
Yale Department of Computer Science
Box 2158 Yale Station
New Haven, CT 06520
In article <·····················@netcom.com> ·····@netcom.com (John Nagle) writes:
>·············@cs.yale.edu (Sean Philip Engelson) writes:
>> Now Available
>> ARS MAGNA
>> The Abstract Robot Simulator
>>
>>... Lately attention has turned to planning for more realistic
>>domains in which micro-world simplifying assumptions do not hold.
>>simple enough to support controlled experimentation.
>>ARS MAGNA's environment and robot models are based on
>>current robotics research, so that the domain is reasonably realistic.
>>At the same time, we abstracted away from many (though not all)
>>real-world details of kinematics and motor control.
>
> Despite being called a "robot simulator", this is a world modeller
>without much reality. It's a 2D grid world, and it is basically
>predicate-oriented,
>not geometry-oriented. The robot is modelled as a point with "sensors",
>"effectors", and "internal carrying capacity", in a world with "walls"
>and "things". You code a planning program in "nisp" (a LISP dialect)
>which does such things as calling "hand-grasp-desig", commanding a
>designated hand to grasp an object. This succeeds or fails, depending
>on the contents of nearby cells, and sometimes fails randomly, so the
>planner has to be able to recover from errors. That's the lowest level
>of detail supported. No kinematics, no dynamics, no gravity, no inertia,
>no friction, etc.
>
>
> John Nagle
Is there a version of this code available that is non-Windowed, i.e. semi-
portable to other systems?
--
Steven Eric Schwartz "It's not safe out here. It's wonderous, with
(········@cps.msu.edu) treasures to satiate desires both subtle and
125 East Wilson Hall, MSU gross. But it's not for the timid."
East Lansing, Michigan 48825 - Q, "Q Who"