Saturday, February 12, 2011

RoboEarth - coloring inside the lines

Full disclosure: I'm not a scientist by profession, nor in the robotics business. I do however have an interest in Artificial Intelligence.

With that said, I see the news about RoboEarth this week. They have a lot of pdf files to read if you're interested. They seem to be mostly top down discussions of how robots in strictly confined domains can cooperate. Think of robots working together in the same building. The word earth in the name seems a bit misleading. It seems more of a shared controller rather than anything like artificial intelligence or IMO even a way to make sophisticated robots cheaper.

It may well work in well defined domains but I don't think that it will scale at all when different domains or radically different robots are mixed together. It all sounds good at a 35,000 foot level, but getting robots to talk the same language is a bit more difficult than it would at first appear.

If we simply look at how humans communicate it is easy to demonstrate why a simplistic view of communications is silly... at best. Say you are driving a car and one of the passengers says 'hey, what is that thing over there?'

Unless you are in a desolate place where there is only ONE thing to look at aside from desolate flat earth, you will have to interpret what in the world the other person is talking about, if that is even possible. It gets even better: Suppose the passenger says 'what's that brown square thing' while you are driving through a city? Communications are difficult, even if you agree what words to use. Do you think I'm joking? Here is a task for you: using any dictionary you choose, describe how someone should tie their shoes without using any pictures, words only. When you are done, get someone to try them out and see how that works out for you.

Take this paragraph about Environments:
""Action recipes may rely on information about the environment, where actions
take place. Therefore, representing environments is part of the specification
of the RoboEarth language. It comprises describing the poses of objects and
pieces of furniture, but also learned knowledge such as the default location of
objects. Using symbolic representations for maps is not suited for uncertain
and dynamically changing environments, so that in this case the model of
the map and all associated sensory data will be saved in its binary form. In
order to create awareness of a map, its meta-data has to be detailed within
the RoboEarth language.""

I don't think this group has fully thought this through, especially since their system relies on the centralized part to make assumptions and decisions independent of the actual robot.

In a limited and well defined domain with strict compliance to format by the robots, yes this could be useful. It's not an answer to any real problems though. It's an answer looking for a question as far as the real world is concerned.

Just my thoughts... too much hype, not enough substance.

No comments: