these interactions. The student demonstrating one system did not know the terms "forward chaining" and "backward chaining." The lack of interaction structure and the independent operation of each modelon are deliberate; Prof. Okino calls the approach "bionic manufacturing." The long-term objective is to create manufacturing systems that are self-modifying and selfgoverning regardless of how complex they become.

The applications demonstrated included robot grasp planning and hidden line removal. In the former, the robot, the object to be grasped, and the gripper each are separate modelons seeking to find graspable faces. In the hidden line removal problem, each modelon is a solid that intersects the others, and the submodelons tell lines how to decide if and how much they are hidden.

Direct Support for Designers

At Prof. Fumihiko Kimura's laboratory at the University of Tokyo, several varieties of CAD are being pursued. Some of their recent work on solid modeling and constraint-based design has been overtaken by new releases in the commercial world, such as SDRC's Level VI, but other work is farther ahead and will yield practical results soon. These include methods for predicting configurations of assembled parts, taking tolerances and imperfect geometry into account, and design of sheet metal parts where only part of the design is given explicitly by the designer.

The sheet metal design project is interesting because it attacks an aspect of design mentioned above, namely, operating with partial information. In this case, flat sheet metal parts must obey requirements, such as having holes in certain places. However, portions of the part, especially of the perimeter, are unspecified in detail. It is known that other parts will intrude at some

places and that all the given holes and Management Methods
slots must be included within the part's
boundary. The computer then suggests
a perimeter shape, which the user can
modify.

Another interesting area is called
top-down design. This is similar in broad
spirit to Prof. Tomiyama's work but is
more focused, less general. The idea is
to provide the designer with geometric
features that have engineering knowl-
edge attached to them. These features
often come in pairs that operate together
but normally belong to different parts
(bearing and seat, screw and hole). The
required knowledge is actually shared
in the pair and ought not be separated
out to the single parts. Some precalcu-
lations are also represented. For exam-
ple, if the designer specifies the load on
the shaft, the correct size bearing is
recommended.

Prof. Kimura recognizes that the main need is to transform CAD from something that produces a model of a drawing to something that produces a model of a product. The exact composition of a "product model" is unclear at this point, but it obviously needs information of both an engineering and a business nature. How many units will be sold in Brazil 2 years from now must be known to the product and process engineers for a variety of reasons. The international STEP/PDES effort is a step in this direction.

At the moment, no top-down approach for creating such models exists. The current approaches are via features. Stringing features together is distinctly bottom-up and could become a shapeless mass unless some topdown structuring is applied first. Yet the top-down structure must be flexible and capable of being revised by the designer without destroying the lower levels. Early attempts at this will likely produce methods that impose a design methodology on the designer. If it is the least bit restrictive or awkward, it will be rejected immediately.

Prof. Fujimoto's research on design practices in the world auto industry has been discussed above. Generally, this follows a common business school research paradigm called "best practices." The goal of such research is not to work out new inventory control algorithms or accounting formulae but to determine what the best companies do and how it differs from what less-capable companies do. The research approach involves interviews, questionnaires, and statistical analyses of questionnaire results. One may find, for example, that companies with high model mix, JIT production methods, and democratic management methods are more likely to have high quality and low cost than companies with other management and operating practices.

Prof. Fujimoto is about to launch a new study on how companies deploy assembly automation. Another project will study design and automation in auto companies, semiconductor manufacturers, and precision instrument makers. Each is a rather different industry with different production rates, quality requirements, and processes. In our discussions, I noted that differences in automation penetration in these industries do not depend as much on the attitude of managers as they do on production rate and the degree to which the processes are understood.

The questionnaire method has revealed some penetrating information that was not widely appreciated outside of the companies themselves. However, it can be difficult to make hard statistical analyses because the method does not admit the usual checks and balances, namely, control sets and double-blind techniques.

Role of the IMS and Other
Government Activities

The IMS was originally proposed by Prof. Hiroyuki Yoshikawa of the

WHAT WILL HAPPEN NEXT

I believe that we are on the threshold of a major increase in the capability of CAD/CAM/CAE, and my Japanese academic contacts agree. The stage has been set for implementation of firstlevel feature-based design. Once a few applications of this come into use, people will see the real potential and demand will grow rapidly. There are two elements to this potential: mustering of engineering knowledge and redefinition of the user-computer interface.

Routine knowledge will be the first to be captured, such as catalog information discussed above. Second will be procedures that experts follow, initially without any deep background other than mimicry, later with some logic branching and case-based methods. The major output from such computer applications that will differentiate them from all past applications will be the first data models of products, in contrast to today's models of the drawing on the computer screen.

These data models will provide significant new capabilities linking product function design to fabrication and assembly process design. Even a little data on product topology defined by feature-connections have proven powerful in permitting complex assembly process planning to be automated (Ref 5), for example. The right kind of data structure definitions will make it relatively easy to create many new and significant tools of this kind; applications will snowball. Providing users with the ability to create these applications will be especially powerful.

The redefined user-computer interface will make computers routinely used for complex engineering, in contrast to today's use for complex drawing. The kinds of information that can be linked will broaden to include some basic process engineering at the functional design level.

However, process engineering in general still takes second place to function engineering on industry's priority list. The potential for redressing this exists in several areas, but neither the companies nor the universities have pressed the issue hard enough. The companies all offer the same explanation, namely, that they have experienced people who can do that now. But they could have said that 20 years ago about ordinary CAD. In other words, the potential is huge, especially if it is joined to functional design to produce true concurrent engineering. Since many Japanese companies have newly launched projects to improve design methods on top of their current capabilities, it is likely that application of computers in these projects will increase and will be extremely effective.

before current research on qualitative methods can challenge existing models. In the meantime, people will always be able to do better, faster.

The result is that researchers in many areas will have to choose what roles they think people and computers, respectively, should have in future design systems. One approach is to try to capture deep knowledge so that the computer can in many ways become the designer. The other is to acknowledge that such capture may be impossible for a variety of reasons. Then one would focus on aiding the designer in doing things that he/she could do in principle but should not waste time on or may not do accurately enough. Examples include sorting, matching, enumerating, searching, optimizing, maintaining constraints and enforcing rules, drawing evocative pictures, and otherwise empowering people to apply their deep knowledge.

Design process management, infor-
mation flow analysis, and design pro-
cess improvement are just starting to
be recognized as subjects for research
and technology. Rapid progress can be
expected in these areas because the
main issues are not hard to model and
several existing approaches are waiting
to be applied. More generally, the
potential for joining engineering and ACKNOWLEDGMENTS
management methods in unified com-
puter models of design is large.

The advantage of the latter approach is that practically every research result will be immediately applicable, and verification of the underlying methodology appears practical. Companies will thus tend to trust and adopt the methods.

I would like to thank the following organizations and people for their help, support, and hospitality. They worked hard for me and made my stay enjoyable and educational:

Geometric dimensioning and toler-
ancing is one of many functional and
process design areas that resists major
improvement because there exist as yet
no firm mathematical models of many
of the geometric variations that have Office of Naval Research:
been used loosely in the past. The advent
of solid models brought these short-
comings to light and international
committees are working on them, but a
complete model may take some years

to create.

The most long range research involves trying to capture deep knowledge in new ways. We already have very sophisticated mathematical models of some phenomena that permit impressive simulations, such as crash and skid tests of cars. It will be a long time

Dr. Arthur Diness

CDR John Sheridan
Dr. Donald Polk
Ms. Roberta Perry

Charles Stark Draper Laboratory:

Dr. David Burke
Mr. Richard Riley
Mr. John Barry
Ms. Judith Repetti