It seems that the predominant current research subjects relating to production technology concern computer networks, notably manufacturing automated processes (MAP). However, considering that things are actually produced on the shop floor, prospects for factory automation will undoubtedly be discouraging if due consideration is not given to the automation of operations performed on the shop floor. It is most likely that growing importance will be placed on the automation of machines used on the shop floor since automated machine operation is essential to producing better products using a smaller number of operators. The main requirement to translate this production theory into reality is the development of more intelligent numerical controlled machine tools and FMS peripheral equipment, including robots.

Taguchi methods have been extensively discussed, including a recent article in the Scientific Information Bulletin [K.O. Bowman et al., “Statistical quality control technology in Japan," 15(1), 57-73 (1990)]. They are used not only in Japan but in many other countries, too. For example, at this conference a paper by researchers at Auburn University illustrated its use in studying robot process capabilities.

SPECIFIC PAPERS AND SESSIONS

Europeans, both east and west, were very well represented (perhaps associated with the Esprit projects) at this conference, particularly Italians. Expert system tools such as Prolog, LIST, C++, etc. were in common use, and there were discussions of expert subsystem cooperation for several researchers,

including several from the Computer and Automation Institute in Budapest. The Italians, in particular, seemed unimpressed with the Japanese research and felt that their manufacturing technology was at least as advanced. There were several papers on FMS simulation with authors from Europe and Japan. There were also several papers on "humanizing" automated manufacturing, by appropriately injecting human decision into the process at key steps. This was called "bionic" manufacturing and requires excellent access to the knowledge database by the human in the control loop.

From a computer science (CS) point of view, a good paper was presented by A. Inamoto, who presented Mitsubishi's conceptual and architectural ideas. Their view is that computerized industrialized manufacturing (CIM) should focus on integrating heretofore heterogeneous islands of automation into a manufacturing system for the entire company. His talk proposed a specific CIM systemrequirement-specification methodology based on conceptual and architectural descriptions. For a conceptional descripdescriptions. For a conceptional description, he stated that objectives, scope, and basic management cycle should be defined. Only with reference to this conceptual description should functional and information architecture be framed. A network should be designed framed. A network should be designed to interconnect existing systems, and finally a physical architecture needs to be configured to establish the methodology. For example, he defines CIM objectives in terms of four major factors (productivity, quality, cost, and delivery) and defines CIM scope in terms of management system, sales and distribution system, production planning and production control system, design and engineering system, production line operating system, etc. A rather lengthy example suggested that this approach has been implemented, at least at one part of this huge corporation. I liked this paper because it tried to generalize

and abstract the manufacturing process, rather than focusing on one particular product or production line.

Several papers discussed extensions to Petri net methods to express operational functions in sequence control, such as synchronization, interlock, and concurrency. These extensions can provide temporal information and other token attributes that are lacking in traditional methods. An impressive example of this was given by researchers from Kawasaki Heavy Industries in the development of control software for autonomous ground vehicles (AGV). A more academic example was given by Hatono and others from Osaka University, who use hierarchical stochastic Petri nets to encapsulate some of the same temporal data.

There was the usual bevy of papers on robot dynamics and fine motor control and manipulation, all highly detailed. There did not seem to be any country bias in the emphasis of this research. The placement of a peg in a hole seems to be a favorite application example, as there were a number of papers discussing that, primarily in two dimensions. One scientist from the United Stats felt that there was too much emphasis on detailed movement planning on the assumption that input data are accurate, and that a great deal of work needs to be done on sensors, especially for hostile environments such as radioactive, vibrating, underwater, etc. Ohio-based Robotics Research Corp. described a "17 degree of freedom" manipulator that they emphasized is not a simulation. Korean researchers described a rule-based approach to alternator assembly using expert system ideas to produce a precedence diagram for assembly.

As far as I could tell the Mitsubishi paper described above was one of the only ones from an industrial laboratory that described a general framework, rather than a particular application. Almost as high level a view was presented

by Futagami and colleagues from Sharp Corp. They have developed a twodimensional computer-aided design (CAD) system that separates the design database from the geometric constraint mechanism. In other words, the design database stores and manipulates design object knowledge and provides an easy way to develop tools for interactive design. The constraint mechanism evaluates and satisfies geometric constraints and the attributes of geometric entities. This latter is rule based.

On the other hand, a large number of the university researchers tried to abstract their ideas to discussions of knowledge representation (mostly frames) and various reasoning techniques, such as Mandami reasoning. For example, Bossink and Gaalman, from the University of Twente, The Netherlands, described general planning and scheduling manufacturing systems. There were plenty of expert systems developed and various methods for developing knowledge bases. From the United States, a few papers had a particular application in mind but developed their methodology from a general viewpoint. A notable example was the paper titled "The Development of an Intelligent Dimensional Inspection Environment in Manufacturing," by C. Menq from Ohio State. Some papers used an object-oriented approach for rule-based reasoning, for example, for production scheduling.

Of the university papers, one of the most general was by Murayama from Hiroshima University, who has a goal of integrating computer-aided design, engineering, and manufacturing by proposing a real time design operation to perform concurrent processing of the modelling step at the same time the integrated manufacturing is occurring. I think that this is an important approach because it recognizes that information from the manufacturing process is needed to correctly model complicated manufacturing. Nevertheless, there were

no concrete applications of this work presented.

There was one full session on global programming languages. This is an effort to provide a standard to describe the entire computerized manufacturing entire computerized manufacturing process. A great deal of work has apparently been done in this area and there are several ISO standards that have been published. Using one of these programming languages would allow a user to specify factory function as well user to specify factory function as well as the networking between various parts of the manufacturing operation from administration down to physical equipment and would include process planning as well. Several of the papers were in the form of proposals. For example, IBM Japan writes:

Since 1987, our domestic [working group] has continually proposed and promoted international standardization and technical information exchange about factory automation. As a result, we reached the conclusion that a global programming language for describing a data model of cells, machines, and parts and a cell environment for processing the data model were very important for standard activities. We hope that our proposal shows one direction for future system and standardization activities. The advent of standardization will accelerate the establishment of future shop floor production systems and will not impose limitations on the construction of such systems.

Although this seems to be an international activity, all the papers were from Japanese researchers. I was unable to decide to what extent other countries participate in this research, although I recall that the National Institute of Standards and Technology has also been studying this problem.

In the area of visual processing, an interesting paper was by Kitajima, Tokyo University of Agriculture and Technology. Kitajima claims that there is plenty of research on both geometry and topology separately but very little on their integration. He would like to develop intelligent robots with threedimensional (3D) sensors. One step is to generate both shell descriptions and solid model descriptions. He has been studying how to do this from two displaced views of 3D objects. His results are positive but still very primitive.

There were many, many talks about applications. One of the most dramatic was by representatives from Sumitomo Metal Industries. (I visited another Sumitomo facility--see below.) They have developed an expert system tool that they have applied to a steel wire finishing line plant which is approximately 700 by 150 meters in area. The particular expert system they developed has more than 300 rules and was developed and implemented within 10 months. It has real time control that supervises the movement of more than 20,000 coils of steel wire each month.

There were a few robot systems described in detail. A particularly novel idea is that of a collection of robots that each operate independently but know of each other's existence--a "cellular robotic system." This is obviously related to the notion of parallel task execution. Each robot is programmed to perform a particular task. In some research no system-wide centralized mechanism, synchronous clock, or shared memory is assumed. Each robot in such a system makes decisions solely based on its sensed information and its current state. Nevertheless, all the robots have to cooperate to accomplish a prespecified global task. Some very elegant theoretical work on this was presented by Beni and Wang from the University of California at Santa Barbara. They postulate a one-dimensional ring of robots with certain given properties

produced here is used on the U.S. space shuttle. The plant's sales are almost $1B annually. The tour participants were impressed with several aspects of this plant:

(1) The variety of machine tools needed

to keep up with the incredible diversity of products manufactured here.

(2) Although conveyer belts were everywhere and the plant is bright and airy, a very large number of workers (mainly women) seemed to be busy doing very repetitive labor-intensive tasks such as sorting, testing, and packaging. Although one view of this was surprise at how unautomated this facility was, my own thought was that if this plant is as productive as it is claimed to be, their room for improvement is staggering.

Sumitomo Electric

This company is over 90 years old and had sales last year of over $40B. Their primary business remains electric wire and cables, but they are active in all types of connecting materials as well as sintered alloy products such as cutting tools, disk and antilock brakes, etc. For the future they see optoelectronics, optical fibers, and applications as key, as well as materials such as gallium arsenide, synthetic single crystal diamonds, and ceramics. They are also anticipating growth in various electronic and communications systems, traffic control systems, and medical and cable TV (CATV) devices, and they are enthusiastic about producing what they call "Japan's first domestically produced high-performance engineering workstations." These are either older M68000 based or two newer series that are built with custom RISC chips (R3000) with 18 MIPS and 3.0 MFLOPS performance. They all run UNIX System V + 4.3

BSD with special capabilities for Japanese character processing. They run X Windows and have networking capabilities, fast graphics, etc. We had an opportunity to see an area of about 1,500 ft2 where these workstations are put together and tested. Some of the testing procedures were impressive, such testing procedures were impressive, such as subjecting each unit to run-in in extreme heat and cold conditions within a specially designed variable temperature chamber. The pace, loving product care, and activity level in this portion of the factory did not suggest that workstations are sold in large quantity. Sumitomo staff explained that most of Sumitomo staff explained that most of their workstations were customized to user specifications. If price were no object I would like to buy one of these--it would be like buying a fine watch. But, frankly, some aspect of this production process will have to be substantially altered if sales volume is going to increase.

The main interest in visiting this plant was to have a demonstration of their automobile location/navigation system. Several other competing systems have been introduced, and I have seen them on cars in the Tokyo area, but this is the first one that I have been able to see in actual operation and ask technical questions about. Approximately 100,000 cars have been equipped with these devices in Japan. Typically they will be installed in higher priced models such as Nissan's Cedric, Mitsubishi Diamante, etc. Currently, more than 1,000 Nissans are being sold each month with this $2,500/unit option. The system includes a color CRT display that is integrated into the car dashboard. (The CRT can also receive ordinary TV programs.) There is a keyboard unit in the passenger compartboard unit in the passenger compartment either just adjacent to the screen or between the seats. A system unit mounted in the trunk contains one or more CD-ROM disks with map information (540 MB), a display controller, mation (540 MB), a display controller, and a locator. Sensors are mounted on

the front wheels and a geomagnetic field sensor is located at the center of the roof.

The tour participants were placed two or three to a car in a collection of vehicles so equipped and sent off for about a half-hour ride with a driver. The route was more or less random on the streets around the laboratory site, including some that would be classified as "alleys" in the West. By keyboard control it is possible to select one of three scales of display, 1/400,000, 1/100,000, or 1/25,000. At the same time a vehicle position marker is displayed at the center of the screen. As the vehicle moves the marker traverses the map. The vehicle location normally shows at the screen center with the map scrolling underneath it. Two display modes are available, "north up" or "head up." Users can scroll the map manually to see other sections and can enter a destination point if they wish. If this is off the screen an arrow and distance indicator show which way to go, otherwise the destination appears on the current screen. Major hotels, department stores, auto dealers, etc. are stored on the CD, which can display them on the map and also display their addresses and telephone numbers. Destinations can be entered either graphically or by referring to one of these prestored data. Users can also store frequently used destinations such as their home, office, etc. Similarly, the vehicle position changes continuously, but under some conditions it must be set by the user, for example, if the vehicle has been moved by towing or on a ferry. This can be done in several ways, including graphically or by specifying that the vehicle is now at some location (hotel, etc.) already in the database.

Map information is currently available for three metro areas, Tokyo, Nagoya, and Osaka--this region accounts for about 30% of Japan's population and about 70% of the system's unit sales. The finest scale display can only

be used where detailed road network data are available; the map scale changes automatically if the vehicle leaves the fine scale area.

During our ride the system worked perfectly, showing the car's location even on very small streets. Backing up, going through parking lots, etc. seemed not to faze it in the least. The screen is busy even though some fine detail information is automatically removed while the vehicle is moving and many keyboard functions are also inoperative at that time. But such a system would not be permitted presently on the dash in the United States, but the concept makes excellent sense in many urban areas.

Later I had an opportunity to talk to some of the engineers. Prototype software is developed in C, although most of it is eventually rewritten in assembler after it is debugged. Program size is almost 400K. The key questions are how well it works and how reliable it is. Of course, I didn't have access to the real details of their programs and can only guess about hardware, but I was very impressed with the care that had gone into their planning for this device. A few examples follow. A gold CD disk is used for reliability. It is waterproof and is claimed to have more than 10 years of life expectancy. (Knowing how often streets in Tokyo are rebuilt, this seems about five times longer than the data will remain valid.) Map data on the CD-ROM are organized by latitudelongitude mesh coordinates and occupy more than 300 MB. There are over 21,000 meshes. On the disk, data are stored repeatedly for speed of access. (The master database is maintained on a Sumitomo workstation in Osaka.) To prevent excessive heat from ruining the disk and driver, a special voltage regulator is used that will only power the unit if the temperature is within its normal operating range.

Since the vehicle location is not determined by satellite or other external inputs, sophisticated “dead reckoning" techniques need to be used. The data input and program to manage this are in the locator portion of the system. This is basically a 16-bit CPU with floating point coprocessor. The two wheel rotation sensors are in the front wheels because tests showed that they are less susceptible to the distorting effects of traction power on the rear wheel drive vehicles in which the system is presently installed. The locator detects distance by averaging the right and left wheel pulses and detects turns by measuring the difference between rotation on the two wheels. Some slip checking is also included. The magnetic field sensor is on the roof to reduce effects from the vehicle. A rotation test is made to calibrate the roof sensor and eliminate bad data from tunnels and bridges. Each vehicle is initialized by running it for 500 meters in a straight line which can then be used to calibrate tire imbalance to within 0.05%. Various calibration coefficients are stored in an EEPROM in the locator. Every dynamic RAM is temperature tested to insure that it will function at low enough temperatures for practical use. I didn't have an opportunity to ask what would happen if the tires are changed, for example, after a flat.

Where is the vehicle? A first approximation is made comparing tire rotation and magnetic field output, after some Kalman filtering. This usually gives location correct to within 5% (say 500 meters in 10 km). The next step is by "map matching," which is really pattern recognition. If by dead reckoning the vehicle is not on a road, the system must adjust. Some correlation coefficients are computed between the computed location and mapped roads. computed location and mapped roads. The distance error and direction error

between roads and current location are computed in an effort to find the most likely road nearby. All nearby roads are investigated along with their branches. Once the vehicle turns there is a good chance that location information can be obtained much more accurately. In Japanese urban areas there may be more than 400 intersections per km, thus map matching needs to be accurate to under 25 meters at worst. Location information is updated on the screen every second. In use, one may see the vehicle "spot" deviating slightly from a road while driving in a straight line, but it always places itself back on a road after a turn.

The Sumitomo engineers expect that their system will improve in several ways. First, they are actively keeping up to date in pattern matching research to be sure that they are using the best techniques currently available. (As this product is clearly profitable I see substantial opportunities for practical applications of research here.) Second, this year receivers/transmitters are being placed on major streets in Tokyo which will provide reference location data to the onboard car systems. Third, they expect to provide computerized route guidance and traffic information (some of this is already available), integrate phone and fax services, and allow users to make hotel and restaurant information from the database on the CD. The price of this system is bound to decrease not only because of efficiencies of production but also because merchants will scramble to be included in the database--"what is the closest Chinese restaurant to my current location and can I get a table in half an hour?" On the other hand, there are questions about keeping the data up to date (perhaps one will be able to purchase CDs like maps at gasoline stations) and its utility in nonurban environments.