Lenny and the Baby Robots

By Mohinish Shukla
Illustration: Amruta Patil (based on a sketch of LNE by Mohinish Shukla)

The late Dr. Isaac Asimov, bless him, generously bequeathed to us a thrilling 21st century, peopled with many, many different robotic beings, mostly made at the US Robots and Mechanical Men, Inc. (US R&MM). As those familiar with the Asimovian universe know (too well), the robots created by US R&MM have fantastically sophisticated positronic brains that obey the 3 Laws of Robotics: 

1) A robot may not injure a human being, or, through inaction, allow a human being to come to harm
2) A robot must obey orders given it by human beings except when such orders would conflict with the First Law; and 
3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

One of the important personnel at the US R&MM is the formidable chief robopsychologist, Dr. Susan Calvin. A particular delight in the Susan Calvin stories lies in discovering the extent of the implications of the Three Laws and their interpretation. Sometimes, somehow, a robot causes some harm to a human being, and it is the job of Dr. Calvin to understand why. 

In one such story, Susan Calvin finds that, due to an error, a prototype robot from the LNE series, intended for mining boron in the asteroid belt, turns out as a robot-baby. So, while earlier robots in the Asimovian universe stepped off the assembly line ready for their assigned work, Lenny toddles on unsteady feet and knows little more than the Three Laws. 

“Lenny” is, for me, a turning point in the journey from the robot as little more than a fancy calculator to a rational creature seeking entry into the human race (as in the sublime story, “The Bicentennial Man”). Lenny marks the coming of a more human robot; one who starts off ready to learn. Rather like a real human baby. Susan Calvin recognizes the immense potential of such a robot: instead of thinking of robots as highly specialized, fancy tools, one could begin to think of them as rational machines that would grow up in a similar way as human babies do. 

But what is this “way”? Do we understand enough about human babies so that – when, sometime later this century, the US R&MM begins full-scale production of the babybots – they would know what to build?

We are much more familiar with human babies than with robot babies. And what strikes anyone who has spent any period of time with babies is that they seem to absorb information from the environment like sponges. How do they do this? And where do they start from?

More than a century ago, the American psychologist and philosopher William James imagined that when a baby was born, its perceptual world was “one great blooming, buzzing confusion”. From this chaotic mélange, the infant slowly comes to discern its mother’s voice, its father’s face, a warm bottle of milk and the telly. In this view, the mind of the baby is an empty slate, on which experience writes down what the world is like. 
For example, consider a property of the world around us: object occlusion. 

It’s quite plain to you that when a billiards ball disappears down a pocket, it doesn’t literally disappear; it’s still there, but hidden from view. Surprisingly, 2-month-old infants don’t appear to have figured this out. (How, you might wonder, do we know what 2-month-olds have figured out or not?) 

Researchers have found that, when 2-month-olds see a ball moving behind an occluder, they appear to treat each reappearance of the ball as a novel event. It is as if, for the baby, the display switches between two scenes, one containing a ball and one not. But, by 4 months of age, babies seem to know that it is a single display; one in which the ball is merely hidden from view temporarily, and they gaze expectantly at the occluder at the point where they think the ball will reappear. It is as if it takes four months for experience to write down on that blank slate: “objects that disappear behind other objects don’t disappear from the world, but are merely hidden from sight”. 

Several such studies come to the same conclusion: very young infants don’t seem to understand the general concept of objects-behind-objects. As another example, upon seeing an image of a triangle partially occluded by a bar, they seem not to be able to infer the presence of a complete triangle lying behind. Now, here is the funny thing: newborn chicks are perfectly able to make just such an inference! Of course, newborn chicks are able to do lots of things that human newborns cannot; for one, they can walk and run around. 

But take a moment to reflect. Do we think that the ability to walk comes out of experience? Clearly not. It’s not like the human newborn, upon encountering flattish surfaces, gravity, friction, two (semi-)controllable limbs extending below the waist and other similar-looking beings walking around, figures out from scratch that putting this limb so, and then the other one so (controlling for any number of external factors), it can walk. Instead, walking seems to follow its own developmental program, just like facial hair, zits, breasts and those curious feelings for members of a certain gender. 

So, typically, you cannot walk when you are two months of age, and people around you would be very worried if you were not walking by the age of three. 

And so, we come to a second theoretical stance, the innateness hypothesis, according to which, human babies are born knowing a great deal of things, and a great deal of other things come to be known in the process of maturation. So, in the innateness view, walking is built into our genes. Of course, you need some experience. Walking is a delicate balancing act, and like anyone who has tried balancing acts will testify, you need the careful adjustment of the weights and counterweights. Experience provides the data to adjust the system of weights and counterweights to make walking possible. And so, infants are rather like astronauts that get to the moon and need some time to readjust, just as astronauts need some experience to fix their moon-specific system of weights/counterweights.

So in the innateness view, we understand occlusion not because we learn about occlusion by observing the world around us, but because, in the course of its development, the baby brain grows little bits that encode the whole concept of occlusion. In fact, from a logical standpoint, how on earth would one ever learn about object occlusion from vision alone? An object can be (partially or completely) occluded, or not occluded. How do we know that the occluded object has not gone for a walk in the park? And notice, you can’t say that you can just crawl around the occluder and check – as soon as you crawl around, the object is not occluded anymore, and the mystery of what happens to the occluded object remains unsolved. 

Why do we believe so strongly that the occluded object remains there? Why do we not imagine that it has gone into another universe? The innateness hypothesis suggests that it is because this is the way we are built. After all, occlusion is an idea. It makes no sense to be able to actually see occlusion. It is an idea that is, fundamentally, inside our minds.

So, from the viewpoint of the US R&MM, the babybot will need to have scores of in-built notions. We do not as yet have a complete list of what in-built notions exist in the human baby, and this is a hot area of current research. We know, for example, that babies seem to come equipped with the capacity of understanding a few small numbers, faces, the motion of biological objects (as opposed to cars) and that part of their acoustic input that corresponds to spoken language. We need to understand what the innately specified mental toolkit of the baby is, before we can endow the toolkit to the babybot.

Current technical advancements in the field of robotics and artificial intelligence are geared primarily towards understanding how to make a babybot that learns from its environment. In fact, a prototype babybot has already been built in the lab. But, here is a thought. Clearly, the babybot is pre-programmed to learn. Now the question is: what is the pre-programming in the human baby that causes it to learn from its environment? In other words, is there yet another component of conceptual toolkit that is specialized for learning just about anything? Well, that is a whole other story. 

What is a bit of a pity is that the human innate endowment, unlike that of Lenny the babybot, does not enforce the Three Laws, which, in the words of Dr. Susan Calvin, set robots apart from humans and make robots “…essentially decent”. n

© mindfields 2007