Robots: Poor Students of Humanity : Scientists Having Trouble Teaching Artificial Intelligence
STANFORD — The effort to build machines that can take over human functions has met with success in some areas, but even objectives that would seem to be relatively basic are proving extremely difficult to achieve, a symposium at Stanford University was told last week.
Robert H. Cannon Jr., chairman of the steering committee for the university’s Institute for Manufacturing and Automation, said duplicating the human body’s facility to do many things at one time is proving a real challenge.
Cannon noted, for example, that many factories now use robots to carry out repetitious chores, but the current generation of industrial robots is restricted to rigid structures that offer only limited control. Simply duplicating the flexibility and control of the human arm has proven a very elusive goal.
The Institute for Manufacturing and Automation, which is sponsored by the high-tech industries of the nearby Silicon Valley, is trying to learn how to build robots that can make decisions and relieve human drudgery, something that has come to be known as artificial intelligence.
He said part of the problem is that people take for granted such things as dexterity, feeling, even vision. Yet those abilities, which include such fundamental functions as reaching out and picking up an unfamiliar object and placing it someplace else without damaging it, are not readily transferable to machines.
Part of the research at Stanford centers on shifting the control for robotic arms from the “shoulders” to the “fingertips.” A human arm, held rigid and controlled only by moving the shoulders, is a very clumsy instrument. By using sensors at the tip of flexible robotic arms, Cannon hopes to come closer to emulating the flexibility of the human arm.
“This turns out, for fundamental stability reasons, to be very hard to do,” Cannon said.
One instrument built by Cannon and his assistants uses tiny infrared transmitters at the tip of a flexible arm. The transmitters emit a beam of light that is “read” by a microchip on the ceiling as the tip of the arm moves. The beam travels across the microchip, causing voltage to vary depending on the position of the tip. That variation, when fed into a computer, makes it possible to move the robotic arm by directing the tip to a specific spot, thus creating far more flexibility than is possible with a rigid structure.
Recognizing that two hands usually are better than one, Cannon hopes it will someday be possible to use many robotic arms in confined, less structured areas.
“How would you feel if you walked into your factory one day and all your employees were working with one hand tied behind them?” Cannon asked representatives of several industries gathered for the symposium.
Larry Pfeffer, also of Stanford, added two other conditions of the current state of robotics to Cannon’s analogy.
“And what if each of your employees were blindfolded and wearing a boxing glove?” Pfeffer said. “It wouldn’t be very easy to assemble a fuel pump.”
The thought of using two arms simultaneously is taken for granted by humans, but it is far more difficult to do with machines, Cannon said. He illustrated the problem by supporting a pencil between the forefingers of his left and right hands. He moved the pencil back and forth, noting, “It’s easy. You can do it yourself. But now try it with someone else” holding one end of the pencil.
Each arm, in that case, must sense the force being exerted by the other.
If robots can be made to “see,” so they recognize objects, and “feel,” so they can distinguish texture, they also must be able to keep from banging into each other, turning the work place into an unpeopled war room.
The term for that, appropriately, is “collision avoidance,” and last week Stanford tested a mobile robot designed to dodge around obstructions. The wheeled robot uses sonar to detect obstructions, much the same as ships use sonar to map the bottom of the ocean. When sonar waves tell the robot’s computer that something is in the way, the computer changes the course of the robot to allow it to inch around the obstruction.
The first test, at about 3 a.m. one morning, was successful, but the robot moved very, very slowly around a chair that had been placed in its path.
“Maybe we tried it too early in the morning,” quipped Oussama Kahtib, a computer scientist with the institute.
Kahtib, a native of Syria who studied for 10 years in France before joining Stanford, is also trying to “teach” robots how to control their own strength. Fingers at the tip of a robotic arm that cannot exercise “force control” could destroy sensitive components by simply gripping them too tightly, Kahtib said in an interview.
“That is very important in terms of assembly line operations,” he said.
Much of the research at Stanford involves using tiny spring-backed sensors to measure resistance, giving the robot’s “brain” some idea of the texture and strength of the object.
But instruments such as that would be very vulnerable to damage, and the “brain” must be able to tell which sensors have been damaged and should be ignored, just as a human brain “ignores the information it receives from a part of the finger that has been scarred,” said Lyn Bowman, an electrical engineer with the institute.
Bowman, who is trying to develop equipment that will enable robots “to tell if something is slipping in the grip,” said that kind of sensory equipment has to respond so rapidly that no “thought processes” are required.
Bowman grew up in Canada, and he noted that when he slipped on the ice as a child, his body compensated for the slippage before he knew what had happened.
“Tactile arrays, like those in the human spinal column, respond to slipping on the ice without thinking about it,” he said. That same ability will be required in robots if they are to have the kind of role many envision in the next generation of high-tech factories.
There is a lot of pressure for progress in robotics, because many futuristic projects call for considerable movement in that area if the projects are to be realized. For instance, the National Aeronautics and Space Administration expects to use robotics extensively in the construction and maintenance of a proposed orbiting space station.
For now, however, scientists and engineers are more down to earth in their goals. With enough progress, Stanford’s Pfeffer quipped, “We might eventually even be able to change a tire.”
