Robots as blank slates for study
by Bonnie Lee La Madeleine
Intelligence and stupidity are delicate topics. They reflect assessments of character that can scar or heal us. Yet, what it means to be intelligent or stupid is, well, confusing. Worse yet, use of intelligence and stupidity as labels are far from universal. Anyone who has ever defended a beloved pet's maligned intelligence would know how fine the line dividing the haves and the have-nots truly is. What makes a fox cleverer than an ant or a gorilla stupider than a chimpanzee? And, even more puzzling, why are robots and androids presented as idealized forms of intelligence?
There are several definitions of intelligence. Some claim that intelligence is reflected in the ability to learn from stimuli. Others say that intelligence lies in the ability to understand, alter and create abstract concepts. Generally, intelligence is assigned to higher cognitive functions once attributed to humans alone. Later, humans granted dogs the capacity for intelligence, later all mammals, and so on. In fact, examples of observable intelligence are more ubiquitous than most might want to imagine. Fruit flies, ants, and the common garden slug can learn. More astonishing is that even those single-celled organisms that many of us tortured in junior high school, the paramecia, are trainable.
Of course, whether the paramecia are truly intelligent depends on the definition of intelligence. However, all forms of life seem to have an innate learning capacity, which may even precede the emergence of a nervous system. This makes it hard to define intelligence in biological terms, as all animal models can learn. Researchers need an empty-headed model for comparisons yielding insight into biological intelligence. For this they turn to robots.
"I wanted to know what intelligence was, to understand humans," explains Dr. Emilia Barakova, a Bulgarian-born scientist with RIKEN Brain Science Institute in central Japan. "I was drawn to robots because they are stupid." Each time she talked about what a robot could and could not do, she chuckled, as you might while watching a cat fall off the sofa. At least the cat would appear embarrassed; a robot would just sit there expressionless. Before it can explore an environment or complete a task, it requires instructions for execution of basic movements, decision-making and map-making.
"How can we teach intelligent behavior," Barakova asks, "if we do not understand intelligence ourselves?"
Picture "Opportunity," one of the two almost adorable rovers on Mars, stuck in the sand not too long ago. Sure it could take pictures that gave the brains back home some idea of the problem, but it had to wait for instructions teaching it how to adapt its tools to digging itself out.
Dynamic behavior that would enable real-time responses to novel situations and truly independent behavior is still not possible in robots designed to be autonomous. Not only would they need to learn how to map new environments and locate themselves within that map, they would also need to be able to recognize the area despite changes to it or their starting positions on subsequent visits.
Until recently, Barakova was a robotics researcher at a German-managed laboratory in Kitakyushu, where she worked on designs to improve automatic mobility and self-navigation in small carlike robots. She joined the lab, which combined the Germans' innovative work with the Japanese focus on humanized robots, from a desire to explore ways to design robots that could learn, think, and be creative.
The problems with full autonomy
Are "Spirit" and "Opportunity" fully autonomous robots on Mars or merely fancy remote-controlled cars? The latter: they are costly, old-fashioned toys with some sophisticated equip-ment. NASA might prefer describing the pair as robotic extensions of ourselves that enable exploration of remote places, but the point is clear.
"These robots are semi-autonomous. In many cases they must be operated from Earth," says Barakova. The rovers on Mars are shining examples of classic robotic design, which develops robust, highly functional, reprogrammable robots.
Barakova herself designed soccer-playing robots for the Robotic Football Challenge.
They had one purpose: to score a goal on a penalty kick. Period. Her robots cannot plan adaptively. Nor do they have a sense of reward that makes scoring so important. This is to say, like "Spirit" and "Opportunity," they are further examples of classic robot design. Barakova recognizes feelingly the yawing gap between robotics and human behavior.
For autonomous movement, the robot needs to localize and simul-taneously create a map or have algorithms enabling navigation of its environment. It also needs to adapt in response to changes in its environment and in its perceptual inputs -- that is, a robot needs to process information dynamically in real-time.
Examining the issues that arise with mobility is only one approach to developing robots with a degree of autonomy. The algorithms involved do enable some multitasking abilities. Barakova's robots were able to complete their tasks without instructions from an external brain, but they neither selected their manner of approach nor realized the goal could be achieved in several ways. They were simply preprogrammed, albeit, not remotely controlled, robots.
And that was the problem for Barakova, whose interest lay with understanding the basis for human intelligence rather than with creating robots born of science fiction. In fact, Star Wars' humanlike androids were so implausible that she was more de-motivated than inspired. This is perhaps why she can be so critical of progress in robotics and still be amused by the fact that sci-fi fans and techno-geeks like me are a bit insulted that she finds robots so stupid.
"For all the novelty in robot shape and engineering, robotics innovation is running in circles," Barakova laments. Yes, robots are now capable of moving in all directions, but robotics labs were doing similar things, playing with the same techniques and tricks, at the start of industrial robots in the 1930s. Robotics development is costly and risky, making researchers unwilling to take big steps in design or concept because of the high costs or the fear of losing funding.
"If any part of a robot's physical design changes, the whole robot must be redesigned from the beginning," says Barakova. "Neither the robots nor the science of robotics appears to be evolving."
Barakova compares the current states of robotics and robot learning to a spiral. A spiral can spin, and grow, but that growth is limited. It can only expand in two directions and can neither incorporate modifications nor dismantle key parts and insert modifications. If a change is made to the structure of a robot, the code must be replaced; it is not flexible enough to adapt its mechanisms to any modification. Current robotics, like the spiral, is simply too rigid. It is unable to integrate that learning into itself. Barakova's goal is to find a way to model behaviors she observes in her children, who learn by observing and then imitating. This is how intelligence emerges. Emergent properties of intelligence are dynamic, and computer models and algorithms seem to suggest ways in which emergence is plausible, but how it works, biologically or robotically, is still a mystery.
Some argue that sophisticated robots with true multitasking abilities can be costly to build, less efficient or require long periods of training. Barakova disagrees, because currently available robots are simply programmed with intelligence principles or brain models. But, she admits, developing robots that fulfill a societal need, as helpers of the ill or elderly people, will be harder to accomplish.
"I want to design dynamically learning robots that can take inputs, learn from them and modify them based on experience," says Barakova. "The process would be more like a mobius strip than a spiral that allows the robot to take new learning and turn it back on current processes. Along the way the whole entity becomes capable of changing itself and being changed by its environment."
Dynamic robots, or robots that can engage, change and be changed by their environments, and bio-mimetic robots, robots designed according to known biological principles, are the next generation of robots. But practice and theory are still far apart. The bio-mimetic and neurorobots creeping out of Japanese labs imitate biological behavior. This is a logical first step. Next would be real brain mechanisms that enable thinking, not mimicry, of behavior.
Success here would lift robotics out of a rut and create an estimated ¥4 trillion market in Japan by 2025, according to the Japan Robotics Association. Design of intelligent robots is the goal of most of Japan's ninety plus robot laboratories, many of which seek to develop humanoid robots that could assist humans in daily tasks. More dynamic, humanoid robots have captured researchers' imaginations.
Barakova acknowledges that a truly independent, learning robot is a long way off, but that is not really her goal. She hopes the crosspollination of brain science and robotics will generate fresh ideas to accelerate progress in both fields and illuminate the nature of intelligence -- how human intelligence distinguishes us and why robots are so hard to build. For sometime yet we will still need the remote control to get our robot to defend our pet's honor from insults to its intelligence, but at least we will know if our pet's honor is defensible. JI