The transformer that provides electricity to the AMS building in Providence went down on Sunday, April 22. The restoration of our email, website, AMS Bookstore and other systems is almost complete. We are currently running on a generator but overnight a new transformer should be hooked up and (fingers crossed) we should be fine by 8:00 (EDT) Wednesday morning. This issue has affected selected phones, which should be repaired by the end of today. No email was lost, although the accumulated messages are only just now being delivered so you should expect some delay.
Thanks for your patience.
from the Notices, May/June 1992 p. 404-407
Army Research Office Holds Workshop to Define Research Directions
Is manufacturing a fundamentally messy endeavor, with dies forcing metal into a crude approximation of the ideal shape, "house of cards" scheduling that collapses when delivery of the right bolts is delayed, and cost considerations dictating quick and dirty solutions instead of careful planning? Or is manufacturing a futuristic marvel, with visual sensors that examine products for defects, robots transporting materials quickly and efficiently, engineers quickly and cheaply cycling through numerous prototypes before hitting on the optimal design? At a recent workshop on Scientific Issues in Intelligent Manufacturing, the answer seemed to be, a little of both.
The workshop, organized by the Mathematics and Computer Science Division of the Army Research Office (ARO), was held March 7--8, 1992 at ARO headquarters in Research Triangle Park, North Carolina. A cargo train rumbled past the skinny pines just outside the window of the ARO conference room, providing a real-world counterpoint to much of the high-tech talk at the workshop: there are a lot of great ideas for making the world more efficient, but old technology dies hard, sometimes for good reason, and sometimes for no reason at all.
Why is the Army interested in manufacturing? Isn't there a whole defense industry set up to cater to its needs, from Patriot missiles to $1000 toilet seats? Not for much longer, says Jagdish Chandra, director of ARO's Mathematics and Computer Science Division and organizer of the workshop. In addition to responding to a government-wide initiative in manufacturing, Chandra explains that, in the future, it won't be economically feasible for the Department of Defense (DoD) to rely on an industry that's dedicated to DoD needs and shielded from competition. As a result, he says, the DoD must support a research base that will contribute to efficient, flexible manufacturing so that other companies will be able to manufacture the products the DoD needs. In addition, the ARO has a mission to support the science and technology research base. He notes that mathematics and computer science provide the fundamental tools in such areas as simulation, prototyping, and control, so his division has an interest in supporting this kind of research. The purpose of the workshop, he says, is to identify scientific issues in mathematics and computer science that will help to solve manufacturing problems, and especially to pinpoint "gaps" in knowledge that need to be filled before successful knowledge transfer can be made.
The workshop brought together about thirty-five participants, mostly from departments of industrial or electrical engineering or statistics, though several were from mathematics or applied mathematics departments and a few from industry. Emerging from the presentations were a variety of intriguing problems that tax scientific and technical abilities, and a degree of disagreement on how to get at the underlying scientific problems that manufacturing poses, on whether those problems can be solved in a cost-efficient way, and on whether the scientific knowledge and the right mechanisms are in place to get the problems solved.
Forming the backdrop for the workshop was the declining ability of the U.S. to compete in global markets. Part of the problem is that other countries can manufacture better products at lower cost and can introduce improvements more rapidly. In addition, points out Anil Nerode of the Mathematical Sciences Institute at Cornell University, many companies, such as General Electric, Honeywell, and Boeing, are cutting back on their long-term research staffs. Industry and government do not recognize the importance of both basic and applied research to manufacturing, he says.
Other problems haunt U.S. industry as well. Ideas for new industrial processes sometimes never get put into practice because of uncertainty over whether the Environmental Protection Agency (EPA) would approve them, says Nerode. Countries like Taiwan, Malaysia, and the Philippines can do things that American industries cannot because those countries have fewer environmental restrictions. Nerode doesn't advocate neglecting environmental concerns, but he raised questions about whether some EPA regulations really make sense. One famous example occurred when New York City was forbidden by the EPA to put its waste into the land, the water, or the air, so it put its trash on barges that floated out at sea for a number of years. In the future, he noted, it will be crucial for industry to not just model its products and processes, but its waste as well.
Another problem is that academia---which is where much of the nation's basic and applied research is done---often doesn't understand the needs of industry. James Solberg is in the school of industrial engineering at Purdue University and is director of Purdue's Engineering Research Center for Intelligent Manufacturing Systems, funded by the National Science Foundation. Solberg says that, for example, many people in academia make careers out of applying operations research ideas to scheduling problems, but this research is "laughed at" in the manufacturing world because it simply does not apply to the problems industry is trying to solve.
As another example, he pointed to some research he had done in the early 1970s on flexible manufacturing systems---that is, systems that can be easily adapted to manufacture a variety of products. Since then, many other papers have been written that cite his initial work. He made some simplifications in his research, and the later papers became increasingly abstract and farther removed from the original problem. At a recent conference on research in flexible manufacturing, he says, very few of the participants had actually seen a flexible manufacturing system in action. Because of this tendency toward abstraction and distance from real-world problems, flexible manufacturing has become a "dirty word" in the real world of manufacturing, says Solberg, and some important opportunities have been missed.
On the other hand, Solberg points out that industry is not always as open-minded as it could be. He says he's been working on the development of low-cost mobile robots that could move materials around in factories, which he said would save money in some situations. But industry just isn't biting, because they don't see the handling of materials as a problem. Jokingly, Solberg described the reaction this way: "We can buy K-Mart wagons and hire people to move materials, what do we need robots for?"
Despite these problems, there have been some clear success stories in which theoretical research has helped to solve some important industrial problems. Donald McClure of the division of applied mathematics at Brown University described some applications of machine vision to manufacturing problems. Although the power of machine vision was initially oversold, it has become an important tool in many industries. A tough problem for visual inspection of products is the recognition and classification of defects. Defects aren't always easy to identify: is it a dust particle, or is it a scratch? Development of machine vision has not reached a point where it is easily applicable to a wide range of problems, McClure says.
Cheeseborough-Ponds, which manufactures vaseline, has successfully used machine vision in a controlled situation. At the end of the production cycle, jars of vaseline are run by a machine vision system which examines each jar for misalignment of the label and lids that aren't screwed on properly. If a jar is defective, an air jet blows it off the conveyor belt. Because the vision requirements were very specific, this system was not difficult to set up, and it turns out to be far more efficient than having people inspect the jars.
Another example McClure discussed is optical character recognition which, being a two-dimensional problem, is easier to handle than three-dimensional machine vision. Optical scanners that can convert a printed document into a computer text file are nowadays available for personal computers. However, McClure says this problem is fairly easy to solve, since it involves standardized characters printed in black on a white sheet of paper. A more difficult problem is posed by silicon wafer manufacturing, in which an eighteen-character identification code is etched in a standardized font onto each wafer. Various steps in the processing of the wafers can deface this code so that its characters no longer conform to the standardized font. McClure says the aim here is to develop algorithms for processing the visual information that are insensitive to the variation in appearance.
This approach---finding methods that are insensitive to variation---is credited to Genichi Taguchi, a Japanese industrial consultant. Ramon Leon of the statistics department of the University of Tennessee at Knoxville described another example of the "Taguchi method." A Japanese tile manufacturer would stack tiles onto a conveyor that would pass through the oven. Because the oven heat was not uniform, and because of the way the tiles were stacked, some tiles would come out thicker than others. Two obvious---and costly---solutions are to install fans to even out the oven heat and send through fewer tiles at a time. Instead, Taguchi proposed raising the lime content from 1\% to 5\% in the tile clay, which makes the clay less susceptible to variations in temperature. Because lime is cheap, this turned out to be a cost-effective solution. In a number of situations described during the workshop, this idea of finding and adjusting the right control parameters turned out to be an important method in solving a range of manufacturing problems.
Medicine is another area that poses intriguing problems for the development of intelligent systems. Although this was not a focus of the workshop, Shankar Sastry of the electrical engineering and computer science department at the University of California at Berkeley described some work he had done on a robotic device to remove polyps from the stomach or intestine. In endoscopic surgery, the physician inserts, either through the mouth or the anus, a device that feeds visual information to a television screen. Watching the screen, the physician can then use the device to grasp the polyp, cauterize its base, and remove it. Sastry and his coworkers have built a prototype device with a "glove" that gives the physician, in addition to visual information, a sense of feel---an important factor, partly because the polyps move around and are not easy to grasp and partly because whether a polyp feels hard or soft can give an indication of malignancy. To test this device, the physicians Sastry works with set for him a "grape-plucking" test, in which the device had to grasp a particular grape on a bunch of grapes, cauterize its base, and remove the grape without breaking it. The device cleared that hurtle, and Sastry says he expects it to be available within the year.
Some of the participants laid out fairly ambitious research agendas. Sastry presented a position paper on "rapid prototyping," which he believes is the key to developing high quality products quickly and easily. He directed his remarks toward what he called "mechatronic" devices, though his ideas extend to other products as well. Mechatronic is shorthand for eletromechanical---that is, electronic systems that control mechanical devices. One example is the antilock brake mechanism option available on new cars. Even though the mechantronic gadget used in antilock brakes only costs about $150 to manufacture, Sastry says that the antilock brake option typically costs the consumer between $1200 and $1500. It took the automobile companies so long to develop and create prototypes for the mechanism that they must amortize their investment over several years of sales.
Sastry's research agenda aims at a "rapid prototyping environment," a neglected area of manufacturing that falls between basic research and production. The main idea is to provide an interface between a graphics package and a three-dimensional "printer." In this scenario, a designer could use a CAD graphics package to generate a visual design of a mechatronic device. Information about various parts of the device are then fed to a three-dimensional output device that creates a prototype out of machinable wax, and the parts are assembled by robots. At each stage, the designer can examine the parts and and see how they fit together, not just by looking at a representation on a computer screen, but by actually handling the prototype. As Sastry points out, having a three-dimensional object to examine provides immediate intuitive information that a picture on a computer screen simply cannot communicate. All of this must be done quickly and cheaply, so that the designer can modify the design and cycle through a number of prototypes.
Nerode also had some ambitious ideas for manufacturing, though his ideas were of quite a different type. He believes that many of the theoretical structures in computer science could inspire development of analogous structures in manufacturing. For example, a computer manipulates symbols while a manufacturing system manipulates parts; machine language in computer science is analogous to controller language in manufacturing; the output of a computer program is like the product and waste from manufacturing, and so on. Nerode put up a list of perhaps twenty such parallels between the two areas. What is needed, he says, is a manufacturing "Turing machine"---a generic system that could serve as a foundation for modeling manufacturing systems. Of course, manufacturing runs into messy problems in physics, while computer science stays in the neater realm of symbols, but Nerode made a persuasive case for an intellectual basis for the parallels between the two areas.
Roger W. Brockett of the division of applied science at Harvard University brought up some similar issues in the area of motion control. Is there a way to describe motion in terms of motion "primitives?" Can a computer language be developed that would manipulate geometric data in the way that computer languages can now manipulate symbolic data? Agreeing on a generic representation is crucial, he says, because if representations of motion control are not device-independent, then the engineering costs for using them are prohibitive. He pointed to Postscript as an analogy for what could be developed in motion control. In wide use for typesetting and publishing, Postscript is page descriptor language that can be used on any output device, be it a printer or a terminal, that has a Postscript interpreter. Before Postscript, the way a printed document looked depended on what kind of printer was used. Although Postscript is not perfect, Brockett notes, it is simple, portable, and in wide usage. He hopes that something similar can be developed for computer representation of motion.
Such problems are difficult because of their generality, which makes the lack of theoretical results a major obstacle. Sometimes if one's aims are more specific, more sophisticated results are possible. This approach is seen in the work of H. Kazerooni of the department of mechanical engineering at the University of California at Berkeley. Among a number of robotics projects he described, one was particularly intriguing: a robotic "extender" that assists people in lifting objects so that, for example, a 500 pound object might feel like ten or fifteen pounds. Kazerooni showed drawings depicting various designs of this device that could be used for different kinds of lifting---one showed a man using the device to load baggage into the cargo hold of an airplane. Kazerooni says he was motivated in part by the large number of back injuries in various professions because of too much heavy lifting, and he has found enthusiastic interest in industry. Having a human control the extender avoids the complication of programming a computer to control the device.
A late addition to the workshop program, Kazerooni brought up some uncomfortable points that the organizers probably hadn't bargained for. His aim is to build robotic devices to perform a specific task, and he decried the trend toward trying to solve "universal" problems to the exclusion of getting the job done. He says he constantly has to redirect his students to the task at hand, for they automatically drift toward the larger, more theoretical problems. Because his work is practical, task-specific, and easy to understand, he asserts, program officers pass it over in order to fund "sexy" research on universal problems, despite the uncertainty of the return on investment in such theoretical areas. There is value in trying to tackle universal problems, he believes, but too much money is going in that direction, especially given today's tough economic times. As Kazerooni hammered away at his point, uneasy laughter spread around the room, for most of the workshop participants were of a theoretical bent.
Although many seemed to find Kazerooni's views extreme, everyone agreed that there are major cultural obstacles that need to be overcome if theoretical research is going to have an impact on manufacturing. For academics, solving nitty-gritty problems in industry is too often a thankless job: it's not the kind of thing that impresses a tenure committee. For its part, industry often neglects the development of the theoretical research that could insure its future strength. In the end, it appears that these kinds of sociological problems are almost as daunting as the scientific ones.