While robots in 3d games are currently in the headlines and disputed in courts of law [English source] [German source] there are application of those robots that go beyond the current staffing of deserted islands with chatbots in Second Life.
The matters involved here are complex. But let's start with the concept:
What if a robotic avatar in a virtual world could both understand natural language and make predictions about your behavior because it "knows" what you think?
Right, this would be helpful for all kinds of applications, from gaming to military to education.
And the first steps along this path have been made by the Rensselaer Polytechnic Institute who created 'Eddi' a virtual intelligence which is confronted with a task that 4-year-olds fail to accomplish.
The test done is the so called 'false-belief test'. I quote:
"In
a typical real-life version of this test, a child witnesses a series of
events in which Person A places an object (such as a teddy bear) in a
certain location (such as a cabinet). Person A then leaves the room, and during his absence
Person B moves the object to a new location (such as the refrigerator).
The child is then asked to predict where Person A will look for the
object when he gets back.The right answer, of course, is the cabinet,
but children age 4 and under will generally say the refrigerator
because they haven’t yet formed a theory of the mind of others." (RPI News)
Now the researches of RASCALS (Rensselaer Advanced Synthetic Characters Architecture for Living Systems) created both a version if Eddi who fails like a typical 4-year-old and one that can master both the parsing of the language and the logic behind the task.
The researchers are planning to install a 'star Trek type holodeck' where virtual characters can interact with humans in a real environment. While the team needs to cooperate with a heavy duty computing array from IBM at the moment to accomplish this task at the moment we might see similar setups in 5 or 10 years more commonly.
Source: RPI
People keep citing this work! But it does nothing like what the press and the weblogs have been saying.
(a) it may do "as well as a 4-year-old" in a very narrow sense on this one particular task, but it is IN NO SENSE as smart as a 4-year-old in general, nor is it intelligent in any interesting sense. I will bet Euros to donuts that the code that implements this task is very specific to the task, and not part of a more general intelligence software.
(b) no one seems to be mentioning that in fact the "intelligent" version of the robot gives the *wrong answer* to the question in the task. Consider: if you were in a room and someone showed you that the gun was in the red suitcase and the green suitcase was empty, and then you were asked to leave, and then asked to return, and they asked you which of the two now-closed suitcases the gun was in, would you REALY assume that it hadn't been switched while you were out? I wouldn't. Ed's mental model of the person in this situation is so oversimplified as to be non-useful.
This is a cute demo of some interesting academic work, but the claim that Ed is "as intelligent as a 4-year-old", or that it actually has an accurate internal model of a person's belief-formation, is COMPLETELY WRONG.
Sorry, just had to vent. :)
Posted by: Dale Innis | July 04, 2008 at 08:23 AM