Artificial Intelligence

Voice Card  -  Volume 16  -  Paul Card Number 16  -  Wed, Aug 8, 1990 11:23 PM

Artificial Intelligence - 1 1/2 page article on MIT's AI Laboratory {The Boston Globe, 30-Jul-90, p. 29}

[There are two articles: "At MIT, robots march toward independence" and "To hackers, it's all doable." The article includes photographs of robots Genghis, Squirt, Attila and several of the people who create and work with the machines. Here's a majority of the two articles - TT]

Genghis is a 2.2. pound robot that can climb over rough terrain. Squirt is one cubic inch in size and contains a motor, battery, microchip, interfaces and three sensors.

Colin Angle received his bachelor's degree in electrical engineering from MIT in 1989. While still an undergraduate, he constructed Genghis, the first legged robot to apply professor Rodney Brooks' ideas on artificial intelligence. Genghis has 57 behavioral modules, each a simple computer program that receives impulses from sensors, interacts with other programs, then activates the appropriate motors. Now 23, Angle is working on Attila, Genghis' successor, which may end up with 3,000 modules. He never took an AI course before he came into the Mobot Lab.

Traditionally, research in artificial intelligence has been based on the premise that AI in a computer must arise from a model of the human mind, translated into symbols and implanted in the machine. Intelligence, in this view, lies in perceiving the world as humans perceive it, reasoning, then acting.

"Nouvelle AI," as its practitioners sometimes call it, embraces a different definition of intelligence, and it reverses science's approach to creating it. Nouvelle AI denies that a mathematical model of the world, as perceived by humans, must be installed in a computer, and holds that "creatures" that can grasp and choose among ways of accomplishing tasks in real environments are, by definition, intelligent. Instead of trying to simulate the human brain and work down from there to practical applications, lugging along the huge memory and computing power the simulation requires, practitioners of nouvelle AI advocate starting with robots the size of rabbits, with nervous systems more like those of roaches and rat than of humans, and working up.

"Problem-solving behavior, language, expert knowledge and application, and reason, are all rather simple once the essence of being and reacting are available," asserts Rodney A. Brooks, the mathematician at MIT's AI Lab who has taken the lead in elaborating the new approach. "That essence is the ability to move around in a dynamic environment, sensing the surroundings" to a sufficient degree to survive and, in the human case, to reproduce, or, in the robot case, to accomplish the programmed mission. The national Aeronautics and Space Administration's Jet Propulsion Laboratory three times has rejected proposals from the MIT group to develop model robots for space exploration, and there has been a hail of criticism in MIT's direction from other AI shops.

"The descent from human-level AI to 'artificial orthoptera' [insects] is not to be recommended," wrote one scientist reviewing a proposal to the Jet Propulsion Lab. "No doubt these mechanical insects amuse the MIT graduate students."

Advocates of modeling the human brain as a basis for AI have defended that approach stoutly in the face of the MIT challenge. "Any intelligent coupling of perception to action, any planning at a higher level than a reflex, needs a model of the world," wrote Charles E. Thorpe of Carnegie-Mellon University, assessing MIT's heresy. "Robots without models make make great protozoans, or even with much cleverness may produce artificial insects, but I wouldn't trust one to be a chauffeur."

Donald Michie, chief scientist at Turing Labs in Scotland, a leading AI shop, last year dismissed the work as "a Great Leap Backward at MIT's Artificial Insect Lab."

The mobot - for mobile-robot-lab - people are, not surprisingly, undaunted. Brooks keeps the most painful critiques posted on his office door. Angle shrugs: "People have been afraid to build little tiny robots because others would say 'what a nice toy.'"

Professor Patrick H. Winston, director of the MIT AI Laboratory of which the Mobot Lab is a part, says much of the resistance is "what always goes on when someone has an idea that threatens the conventional wisdom. There are always people looking for ways to dismiss it and carry on with the old idea. That is natural in science and engineering."

MIT's AI Lab encompasses 202 faculty, staff and graduate students and last year spent $8.9 million in pursuit of interests ranging from robots to parallel computing to vision. According to a recently completed section for the MIT president's annual report, seven of the 22 professors in the AI Lab are working on aspects of robots, the greatest concentration of interest in the shop.

The idea that complex behavior could emerge from comparatively simple systems was being kicked around in AI circles long before Rod Brooks began to make waves; so was the intuition that AI had gone too far in separating the internal reasoning of its creations from their external capacities to perceive and act.

Theories that complex behavior could simply emerge, without having to be installed by deliberate human effort, were attractive on several counts: For one, they opened up the possibility of creating behavior without having to understand it first, a conceptual advantage that grew as the difficulty of understanding the nature of intelligence grew progressively more daunting. For another, such theories offer a way out to those repelled by the notion of reducing the concept of thinking to mechanical functions that can be replaced by mathematics.

But until Brooks came along, there was no practical approach to building a from-the-bottom-up system. He conceived of layering, adding module upon module of functions: The capacity to wander about, followed by the capacity to explore, followed by the capacity to build internal mathematical maps of where the robot had been. As higher levels are reached, conditions would encourage the machine - or would it be creature by this point? - to make plans for changing its world.

Each individual module is capable of activating the robot, in sharp contrast to systems built on internal models, where stimuli are analyzed and actions chosen in a centralized fashion, and many stimuli may be required to activate the robot. When responses that conflict are activated - for example "hide in the dark" vs. "go to the noise" - a system devised by Brooks to organize the layered processes lets one response go forward and blocks the other.

Described originally in 1986 and refined last year, this system - called submission architecture - has been subject to criticism that it is just a fancy computer program that falls short of artificial intelligence. But also last year, Patti Maes, a Belgian AI researcher then working at the lab, devised an overlay on the subsumption architecture which appears to make possible much more complicated and sophisticated behavior. The approach is analogous to the way hormones work in biological systems. Instead of producing a certain behavior - say, stopping in the face of danger - only in response to a preset level of activation in the sensors for danger, the robot becomes cautious when other sensors detect certain kinds of conditions, and then will stop with less stimulation of the danger sensors.

Think of a household robot setting out to mop the floor. In its background - the Maes overlay - is awareness of a need to recharge itself at some point, analogous to human hunger. The robot decides there is enough charge in its batteries to do the floor before recharging, but signals of the charge level keep circulating, gradually becoming stronger as the charge level decreases.

The children come home from school and continuously redirty the floor; the robot keeps trying to mop it up. But instead of running itself down, the robot eventually changes behavior in response to its power need and puts aside the floor until after a recharge, much as a housekeeper might put aside until after lunch a task that has proven more complicated than expected.

Professor John Laird of Michigan, a leader in tradition AI, says, "You've got to put your scientific bet somewhere, and Rod and I are putting it in different places. But he has systems that do things. AI had sort of ignored that for a while, and it's had a very good affect on the AI community.

"The open question of his research is whether it extends upward to what would be called higher levels of intelligence and problems. That is actually his research agenda. Rod has demonstrated rudimentary planning and learning; it is still an open question whether it will scale up."

Whatever happens with Brooks' work, attempts to model the human mind will continue, Laird and others are certain.

"Modeling of human cognition will help us understand people better," he reasons. "And on the functional, engineering side, well, humans solve problems. We would be crazy to ignore an existing system that is intelligent and can do the things we want our [mechanical] systems to do." Such confidence in the value and richness of the field of artificial intelligence - whatever its evolution, whatever its commercial failures so far - is widespread.

As a noted mathematician explained it recently, "The idea of artificial intelligence was a brilliant and large idea from the beginning. The first people [in the field] invented the idea. The idea produced children and grandchildren, and that was a great accomplishment, even though the grandchildren scorn the grandparents.

"It has both led and exploited the technology of its time."