Robot Dogs Evolve Their Own Language

bab00n writes According to this article at The Engineer Online, researchers led by the Institute of Cognitive Science and Technology in Italy are developing robots that evolve their own language, bypassing the limits of imposing human rule-based communication. The technology, dubbed Embedded and Communicating Agents, has allowed researchers at Sony's Computer Science Laboratory in France to add a new level of intelligence to the AIBO dog. The robot dog has learnt to see a ball and tell another one where the ball is, if it's moving and what colour it is, and the other is capable of recognising it.

Don' think so...

[cp.tar]

"This is not only important from a robotics and AI perspective, it could also help us understand how language systems arise in humans and animals," Nolfi said.

This is all very fine and dandy, but I don't believe that mimicking what is presently known about human language capabilities will help us understand it better.

The technology was, if I understood the article correctly, built on the foundation laid by cognitive science. It mimics chldren's curiosity, it begins from the general semantics (i.e. selecting an entity), goes on to phonology (i.e. the shape of the symbol for the entity), and deals with finer points (morphology, syntax) in the end...

I'd be very interested to see how it goes on, but I really don't think we'll be seeing a huge breakthrough in cognitive science.
NLP, maybe... almost definitely, if we can get machines to learn human languages.
But I really doubt the humans and animals part.

Re:Hmm...

[Bastian]

I wouldn't say it's impossible; a lot of computational linguists have been working on this particular problem for a long time. The Aibo team was pulling from a lot of existing research.

I doubt the kind of language these dogs are using is very similar to any human language. It probably doesn't even have a recursive grammar. Something without that would be a whole lot easier to implement than anything approaching natural language - what they're saying probably resembles a very simple IPC mechanism more than anything else.

Re:Similar stuff done at aibohack.com

[Gulthek]

This isn't about object recognition, that was just an example, but about communication.

Was his Aibo able to teach another Aibo what it knew?

Here's the thing with "A.I"

[insanarchist]

To me, it would seem the most seminal part of creating AI is to somehow instill "wants" and "needs" into machines. Without those, there's really no intelligence. When it comes down to it, the only reason we (humans) do anything is to be happy and to survive; how the hell do we make a machine want/need to be happy/survive? Interesting stuff, to be sure, but we've really got quite a long way to go.

Re:Anyone else feel threatened?

[Daniel Dvorkin]

These dogs are stupid. REALLY stupid. I remember a BBC documentary about AI and robots and they flipped the dog over and put it upside down in their back lawn. The dog tried to get on its feet and after failing several times said something like "I need help".

How is that stupid? Getting up is a pretty complex process; we think it's easy beacuse we do it a lot, but have you ever watched a little kid trying to learn to walk? The "dog" tried to figure out something it couldn't do, realized that it couldn't figure it out on its own, and asked for help -- hell, that's a lot smarter than a lot of humans.

Re:I like this idea

[jtogel]

While computing power is of course good have, I don't think the computing power of individual agents is a major factor hindering development in this type of AI. Our own research (which is similar in spirit to Nolfi's, from TFA) points to the importance of appropriate sensor setups, environments and tasks, and that much can be done with simple neural networks. (On the other hand, if you work in simulation, much processing power might be needed for simulating the environment of the robot.) I recently wrote an article about this, and how computer games might provide the appropriate tasks and environments.

Detailed Explanation (And Why This Is Important)

[goat_roperdillo]

Despite the generated jokes about dogs and the French, and the "oohing and aahing of the crowd at the AIBO robotics soccer games broadcast on U.S. national television, this is not merely "cute". This may be the most important research that you have ever read about.

Researchers Luc Steels and colleagues at Sony's Paris Computer Science Laboratory [csl.sony.fr] in France have performed a series of remarkable experiments demonstrating the development, from naught, of spoken language among robots. Words, grammar and semantics evolve spontaneously among cooperating robotic agents initially programmed with a small base set of ground perceptions and behaviors. And from the development of language arises cooperative group (intelligent) behavior.

Enhanced AIBOs are initially programmed to recognise simple stimuli from their surprisingly limited hardware sensors. Over the course of several hours or days, the AIBOs learn to distinguish objects and how to interact with them. A built-in curiosity system ('metabrain') continually directs the AIBOs to look for new and more challenging tasks and to cease activities that are not fruitful. In time they develop more complex tasks, just as do human children.

Like children, the enhanced Sony AIBOs initially babble ("argue?") until two or more settle on a sound to describe an object or aspect of their environment. Over time the group gradually builds a lexicon and grammatical rules through which to communicate. Agreement on word usage spreads through the population as terms for similar meanings compete for acceptance. For example, the robots develop the language structures to express that a red ball is rolling to the left. Just as human twins sometimes develop a unique language in which only they can communicate, the enhanced AIBOs (which are clone-like and similar to twins) develop their own language.

Language analysis and generation are part of Good Old Fashioned AI (GOFAI) [wikipedia.org] and have been studied extensively for decades by AI researchers. In the past several decades GOFAI was challenged by Nouvelle AI (Situated AI) [usfca.edu] championed by Hans Moravec [cmu.edu] and Rodney Brooks [mit.edu]. This alternative approach holds that true AI will not arise from formal mathematical systems but instead from robotic behaviors which have a subsumption architecture [wikipedia.org] as an overall organising principle for the individual robot. This architecture consists of layers of behavioural modules, each capable of carrying out a complete but simple task. Steels' enhanced AIBOs are embodiments of just such a subsumption architecture and provide strong support for Moravec's and Brooks' hypotheses

Prior to Luc Steels' experiments, no one had experimentally demonstrated how language develops among intelligent agents. Steels' experiments are no less than stunning: in a controlled environment AIBO robots develop their own words and grammars for objects in their environment. All aspects of human language development are mirrored in these experiments: words compete for acceptance in the population, new words are created, and grammatical structures arise spontaneously. Steels' work also addresses the idea of a "robot culture", since it is in the context of a population of cooperating agents that language becomes most useful.

Contrast this with the W3C's Semantic Web [w3.org] effort, which has received much more interest and money in recent years due to the growth of the Internet yet has proven far less fertile. In the Semantic Web there are multiple competing "ontologies" (roughly, data dictionaries wherein all terms are strictly defined by specialists from their

Re:Similar stuff done at aibohack.com

[TinCanFury]

The simple way is to program imitation and have the bot perform the trick when there's any sort of audience.

Now, if the dogs could describe the necessary action instead, I'd be extremely impressed.

though the learning through imitation method may leave them more open to learn and improve upon what they learn.

I don't program, but I've seen other people do things and in my imitation improve on what I learn, so I figure the same would hold true, whereas if you rely directly on being "copied" from one dog to the other, this learning and improving wouldn't occur.

Re:I like this idea

[Intron]

I am kind of worrying about what a group of agents might learn to communicate while wandering around in GTA: San Andreas. Maybe in the interests of society you should limit this to Kirby.

Re:I, Robot

[Memnos]

I do as well, as long as they fetch and refrain from unlimited sloberring. However, frivolous as these pursuits of theirs may seem, remember that very little information is lost in pursuit of software engineering -- it's difficult because it's the chase of entropy's ass. The Japanese, as all other nations at this time, lack the expertise to capitalize on our acheivements -- now. But a day will come when we will incorporate their or someone's discoveries into our own bright ideas, and the reverse. Be forwarned: software is cumulative, and you will one day admire (or fear) what you may have had contempt for before.