Principal barriers to the future of autonomous assembly

Citation

Lowe, G. (2007), "Principal barriers to the future of autonomous assembly", Assembly Automation, Vol. 27 No. 4. https://doi.org/10.1108/aa.2007.03327daa.002

Principal barriers to the future of autonomous assembly

Keywords: Assembly, Robotics, Learning

The missing links in fulfilling the vision of a fully autonomous assembly system, capable of assembling a wide variety of products are – assembly task detection status, the need for manufacturing of smaller assemblies and commercial forsight. The major factor behind this predicament, is to date, no one has recognised that autonomous fault diagnosis closes the expert system self-learning loop, nor has a scientific methodology been devised for solving fault diagnosis.

The most prohibitive barrier to autonomous assembly cells is fault diagnosis. The methodology involves identifying when a fault has occurred and those properties that were responsible for the failure. Just as significant, are the properties and reasons for the success of an assembly task.

Developing an autonomous assembly system is not that difficult. Most of the building blocks exist: self-learning algorithms, expert systems, simulated design and assembly domains capable of compiling into device dependent instructions for robots, assembly sequence ordering models, programmable jig fixturing, design for robotic assembly and programmable part feeders.

For a system to be self-learning, it requires detailed feedback on actions and outcomes. If these attributes can only be achieved through operator input, then the system is no longer self-learning and is therefore superfluous, as the operator would be making a judgment of the system status and ultimately the learning decisions. For a review of the task detection methodology (Lowe, 2006).

Assembly task detection, requires detection of assembly progress and identification of specific factors relating to the failure or success of the operation. It is inadequate to just identify that an assembly has succeeded or failed, an explanation for this follows.

Traditionally, manufacturers have looked for failures because this is the unwanted state, and they have done so by inspecting parts prior to assembly and at completion of assembly. Why? Because these are relatively easy to do and have a low cost associated with them. Any sophisticated testing has a prohibitive price tag and can be readily replaced by an operator, who can test for functionality and make any decisions, repairs or adjustments prior to adding any more value.

In instances where operators make these decisions, they rarely understand the cause of the failure. This has been one of the reasons why failure modes have not been thought about in a fundamental way. Another reason is many researchers have said it is not possible to determine the cause of a failure because failure seldom repeats with exactly the same characteristics (Lopes and Camarinha-Matos, 1995), and have therefore ignored the possibility of finding approaches to solving the failure problem.

Such views indicate a fundamental problem. That is, the methods of assessing the success or failure of an assembly operation focus at too high a level to adequately describe the events. In order to address these inadequacies, a list of part descriptor properties devised by Lowe and Shirinzadeh (2005), focuses on the properties of failure conditions. The properties are measurable, and have relevance to the three phases of an assembly operation. Namely, pre-assembly, concurrent assembly and post-assembly phases. In addition to this, the emphasis is on evaluating what constitutes desired end states.

The complexity of the measurement problem, and the explanation of the reason for failure are very complex problems to solve due to the many variables which are encountered and the difficulty in describing exactly what happened. The dichotomy of the problem resides with the issue that if no knowledge exists as to what to measure, it is then necessary to measure everything. Once every possible measurement is made the problem becomes: “what is relevant and how does this explain the series of events and the final state”. The part descriptor properties address these problems by limiting measurement to key properties known to impact on assembly operations.

Not only do they offer an effective means of assessment, these properties are readily processed in self-learning and expert systems. Providing a means of assessing parts for which the system has no prior knowledge.

This then leads us to the second most prohibitive barrier in the quest for autonomous assembly systems. The viability of autonomous assembly systems, and in particular robotic systems, depends on two commercial decisions. Is it worth investing in the development of assembly state research? And can assemblers continue to rely on cheap labour as a means of mitigating the risks associated with developing an autonomous solution?

There is a trend in which electro-mechanical products are becoming smaller and smaller while increasing in complexity. Our ability to continue to manufacture manually increases in difficulty with this trend (Whitney, 1996). Therefore, the shift of manufacturing to low-labour cost countries has at best, a limited reprieve before investment in technology infrastructure is not an option.

The ability to produce a variety of high-quality products on one autonomous system, opens opportunities to reduce transportation to market and regains control over technology ownership and development, and product quality. When we talk about assembly of a variety of products, these products are often considered in terms of group technology. This is not viewed as a limiting factor to the use of a fully autonomous system.

The ability to produce a variety of high-quality products on one autonomous system, also improves the commercial viability of autonomous systems and reduces commercial risks associated with payback on a single product or product range.

The development of “assembly task detection status” can also be adapted to existing production processes, raising the prospect of line control through expert system feedback. This technology, is possible for dedicated assembly and manual assembly, as well as the principle focus of robotic assembly. Thus, knowledge-based self-learning can be used as a teaching tool and a modelling tool.

Support of research and development of detection tools will open many opportunities for autonomous assembly. Until such steps are made, the majority of assembly tasks will remain as primitive manual tasks. Resulting in missed opportunities for robot vendors, manufacturers of high-tech products, and developed countries trying to retain their principle competitive advantage – knowledge.

Gordon Lowe Honorary Research Fellow in the Faculty of Information Technology, Monash University, Melbourne, Australia. He can be contacted at: gordon.lowe@infotech.monash.edu.au