From: Aaron Davidson (davidson@cs.ualberta.ca)
Date: Fri Dec 24 1999 - 19:48:06 MST
Max M wrote:
>A modeler for input and world building, and a simulator trying out different
>moves in the model before doing them in the real world.
>
>But knowing me and all my other great ideas somebody brighter has probably
>gotten the idea sometimes in the sixties.
>
>Does anybody know about research material among these lines?
Max, you have just described a large portion of what "state of the art"
vision and robotics researchers are working on and have been working on for
years.
Robots with cameras examine a scene, perform edge detection and object
recognition and from that they build a 3D internal model. By simulating the
3D world w/ real time updating, they attempt to plan actions which lead to
the desired goals.
The 3D in Doom/Quake style games is vastly simpler than a real world
problem where fine control is needed. The AI in these games cheats anyhow
-- they usually have information better than what is gleaned from their own
perspectives. They can't be easily compared to real-world tasks.
Currently, the object detection is imperfect and very computationally
intensive, as is searching for a plan in a real 3D world -- the
combinatorial explosion effectively makes the search space infinite.
The most successful systems still take glacial time frames to image, plan,
and act. Apparently one of the biggest obstacles is time-delays between the
real world, the internal model, and the robotic motions. The computations
take too much time and the whole system gets severely out of synch and
flounders.
A common task is to have a robotic arm place a cube through a small square
opening, just large enough for it to fit. This is much like those toys for
toddlers where they have stars and circles and other shapes to fit into a
puzzle with matching openings. This can take the state of the art internal
3D representation systems *hours* to solve. Many other alternative systems
have fared much better. This is not to say that these alternative
techniques are better -- it could be that we just do not yet have enough
juice to power these systems, but when they have what they need they'll
become experts.
However, I feel that this research is a dead end.
I recently saw a talk by a researcher in the field (Dr. Nicola Ferrier from
the University of Wisconsin). She was working on one of these aternative
methods -- a system using edge detection to find predefined objects with
knowledge of various transformations that could be performed (tightly
coupled with the robotic control needed to carry out the transformations)
on the object. It fared a lot better than the traditional approach but was
still, by my standards, very poor. I left her talk with a bitter feeling
that the whole field was quite pathetic.
Rodney Brooks has a novel (and nowadays vastly mimicked) approach with his
'subsumtion architecture' which uses a hierarchy of representationless
control systems and tight sensory/motor feedback loops with the
envirnoment. Unfortunately, this system does not yeild as nicely to
high-level goal planning, but it has worked wonderfully for making little
robotic insects ;-)
I'm not too familiar with other techniques being tried, but I know there
are many. I know neural nets have been used in some systems, but I do not
know what their preformance has been like -- and such systems will be
riddled with all the problems associated with creating and training neural
networks.
I hope these two links are enough to start you off:
Brooks, Rodney. Intelligence Without Representation. Artificial
Intelligence 47 (1991), 139-159.
http://www.cs.ualberta.ca/~ree/cmput366/brooks.ps
http://mechatron.me.wisc.edu/
Dr. Ferrier's research: the site has all sorts of information including
quicktime movies of their robots in action.
Cheers,
Aaron
+-------------------------------------------------------------------------+
| Aaron Davidson <ajd@ualberta.ca> http://ugweb.cs.ualberta.ca/~davidson/ |
+-------------------------------------------------------------------------+
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:06:12 MST