The May issue of Discover has a transcript of a roundtable between the editor in chief, Corey Powell, and four researchers in robotics. It’s an interesting conversation. I found the following quote from Rodney Brooks (founder of iRobot) illuminating:
Rodney, you've talked about four goals that robot researchers should be aiming for. What are they?
Brooks: First, the object-recognition capabilities of a 2-year-old child. You can show a 2-year-old a chair that he's never seen before, and he'll be able to say, "That's a chair." Our computer vision systems are not that good. But if our robots did have that capability, we'd be able to do a whole lot more.
Second, the language capabilities of a 4-year-old child. When you talk to a 4-year-old, you hardly have to dumb down your grammar at all. That is much better than our current speech systems can do.
Third, the manual dexterity of a 6-year-old child. A 6-year-old can tie his shoelaces. A 6-year-old can do every operation that a Chinese worker does in a factory. That level of dexterity, which would require a combination of new sorts of sensors, new sorts of actuators, and new algorithms, will let our robots do a whole lot more in the world.
Fourth, the social understanding of an 8- or 9-year-old child. Eight- or 9-year-olds understand the difference between their knowledge of the world and the knowledge of someone they are interacting with. When showing a robot how to do a task, they know to look at where the eyes of the robot are looking. They also know how to take social cues from the robot.
If we make progress in any of those four directions our robots will get a lot better than they are now.
That’s a clever marketing ploy, I think. It makes things sound a lot simpler to break down the problems into easy (2-year-old) and harder (9-year-old).
But wait a minute. What’s he’s actually saying is, we need robots that work like 9-year-old children!
After all, a 9-year-old comes with the 2-year-old object recognition and the rest already built in.
It’s not like the problems solved by younger children are any easier. The fact that children learn object recognition before mastering grammar doesn’t mean that object recognition is simpler to manage. It may mean that grammatical ability evolved in primates that already could recognize objects. It certainly means that the brain develops in ways that entail learning to recognize objects first – not at all irrational considering the requirements of 2-year-old life. Two-year-olds aren’t going to be teaching much, they don’t need the 9-year-old social awareness. But they do need to recognize objects.
Is the ontogenetic order of these behaviors in children necessary? Or is it an accident of evolution? The answer does impact our choice of strategies for replicating these behaviors in silico. I expect that you do have to recognize objects to be able to understand someone else’s recognition of objects. But do you have to understand language in order to have human social understanding? Some scholars would say yes, others would say these are separate “mental modules” that in principle could occur independently.
Maybe the engineering problem will help us clarify the evolutionary one. It turns out that there was a school of thought devoted to the idea, “Evolutionary developmental robotics.”