Robots, unlike stand-up comics, are not adept at thinking on their feet. Should a heckler interrupt a planned routine, or an irresistible off-the-cuff opportunity arise, the comic can sense these changing situations and adjust. A robot on the factory floor, though, is stuck with its plan amidst a changing landscape, and will keep spot welding, for instance, on a non-existent automobile frame because of a delay on the assembly line.
Now, an engineer at Washington University in St. Louis and two of his former graduate students have stated a theory and devised an algorithm that will make robots as deft and nimble as Robin Williams on "The Tonight Show." Moreover, they have demonstrated their theory and algorithm with robots in the Washington University laboratory.
T.J. Tarn, Ph.D., professor of systems science and mathematics in Washington University's School of Engineering and Applied Science, Mumin Song, Ph.D., technical specialist with Ford Motor Co. in Detroit, and Ning Xi, Ph.D., assistant professor of electrical engineering at Michigan State University, are the first to develop a theory that integrates low-level data sensing, or gathering, with high-level planning and decision-making processes.
The theory, encapsulated in an algorithm, completely automates the process whereby a system must adapt to changing conditions. For years, manufacturers with automated systems have dealt with floor malfunctions in an ad hoc way, involving a lot of human intervention and wasted time and dollars. With the adaptation of the Tarn/Song/Xi theory, manufacturers will be able to let the robot -- or other automated process -- go with or adapt to the flow on its own.
In the case of a robot on the factory floor, the robot would use the all-purpose algorithm to halt work when the assembly line is out of synch and recommence activity once things were in order again. This leaves the plant manager out of the picture, allowing him or her to pursue other tasks. Similarly, in air traffic control, an airplane's automated computer system could relay data through the algorithm to traffic control headquarters, letting headquarters know the plane's location and other parameters to determine the plane's ultimate arrival time. This would lessen the strain of the controller's job.
The Washington University algorithm is called the Max-Plus Algebra Model. It combines task-scheduling, action-planning and control, and it solves a problem that is two decades old.
"The biggest problem in scheduling is to get the machine to fit into the whole system," says Tarn, who is director of Washington University's Center for Robotics and Automation. "Scheduling is on the high-level of control; real-time sensing is on the low level. How to coordinate the two levels automatically has been a wide-open problem for 20 years in the field of intelligent control. We've solved it theoretically with this model, and then we've proven it in the laboratory."
The three were awarded the Best Paper Award at the 1998 Japan-USA Symposium on Flexible Automation, in Otsu, Japan. The paper has been published in the conference proceedings. The work is sponsored by the National Science Foundation.
The Max-Plus model will have applications in automatic control systems far and wide. It can be used in many automated situations where high-level controls must be coordinated with low-level ones. It also could be a boon to the medical industry, where certain surgical procedures, such as artificial hip replacements that require a super-human steady hand, are performed with robotic instruments.
Stop And Think
In Tarn's Washington University laboratory, the engineers programmed a Puma robot to pick up three objects of different height, which spun randomly on a rotating disk. The robot was programmed to pick up the objects in descending order, from tallest to shortest, and then place them in a pre-assigned place. A camera within the robot identified the object by height and instructed the robot on which object to pick up.
The researchers placed an obstacle in the way of the robot, which, as programmed, immediately stopped its activity. When they removed the obstacle, the robot began its task again without human assistance and pursued the exact object it was supposed to choose.
The smooth transition was made possible by the Max-Plus model, which analyzes the real-time disturbances, communicates the problem to the high-level control, and halts the robot and tells it to proceed when the road is clear.
"What the model does is enable the robot to stop and think," explains Song, whose doctoral dissertation was based on the project. "Stopping is the key thing. It gives the machine time to "think" and then feed back data to the upper level."
"In this kind of situation today, an engineer goes into an emergency mode, pushes a button and stops the whole manufacturing process because the robot will keep going as it had been told to, and you have a chaotic situation," Tarn explains. "This is very undesirable. But with our model, the algorithm knows where the robot is in the process, stops the robot and communicates the data to the high-level manager's computer. This way, you don't have to shut down the whole system, which is where the cost problem lies. The algorithm also enables the robot to re-start its task once the problem is corrected."
Many automated systems have computer codes installed that can deal with certain programmed malfunctions. However, the codes are heuristic or rule-based, meaning that they can only deal with known manufacturing errors that have arisen before and have been described mathematically. But, as in the case of the stand-up comic, who knows what an audience is going to throw your way?
"Heuristic code does not begin to exhaust all problems," says Tarn. "This algorithm is all-purpose -- it can deal with any unstructured event. It is getting a good deal of attention from industry."
Editor's note: Photos of T.J. Tarn and a Puma robot are available upon request.