Fuzzy future for robotics?

Citation

Billingsley, J. (1999), "Fuzzy future for robotics?", Industrial Robot, Vol. 26 No. 4. https://doi.org/10.1108/ir.1999.04926daa.002

Fuzzy future for robotics?

Fuzzy future for robotics?

John Billingsley

Keywords Fuzzy logic, Robots, Neural networks, Control

Which way is robotics research going? I spent the first week of April at the conference of the Australian Automation and Robotics Association. Some of the presentations were entirely practical, like the pilotless model aircraft which could fly for days without human intervention, gathering meteorological data. Some were intentional fun, like the presentation on robot soccer.

A self-confessed civil engineer suggested that we should regard the diggers and tracked vehicles in a mining or construction site as an army of collaborating robots. Certainly there are diggers with all the articulation of an assembly robot arm while the haulage vehicles have all the mobility of the mail-serving robot which was featured in another presentation. "All it takes is the strategy and some software".

Although the robot revolution is coming about more slowly than earlier soothsayers predicted, applications are proceeding on lines which are not totally surprising. As it becomes ever easier to add sensors to a process and interconnect the computing power, so manufacturing is becoming "smarter". More of the details of quality control and performance testing can be thought out at the design stage, while solid modeling enables the parts designers to harness more imagination.

These advances do not involve the dramatic emergence of robots as such, but imbue less obvious processes with the essence of robotics, the linking of sensors, actuators and control theory into a functional "intelligent" unit. It is when the researchers strain to push this quality of intelligence artificially forwards that there might be cause for concern.

In recent years, research applications have become increasingly peppered with the words "neural", "fuzzy" and "genetic". Have these been added in the spirit in which chlorophyll used to be added to toothpaste, or is there real substance in the aims of the researchers? The answer has to be, "A bit of both".

If motors and mechanisms are the muscles of robotics and computers their brains, then control theory must be an important part of their souls.

Over the decades, robot designers have been let down badly by linear theory, which does not recognise that motors saturate and bearings have friction. A "stiff" robot axis demands a saturating, nonlinear controller in which integral action has no part. Industry's answer has been to turn to the age-old methods of the PID controller and the PLC. (They don't HAVE to use integral action!)

Fuzzy control tends to revolve around lumping sensor signals into coarse values such as "Positive Large" and "Near Zero". In the process much valuable data is discarded, but the system performance is still seen to be an improvement on non-saturating linear control. Why? Because the method allows the inputs to be driven into saturation when this is necessary for a faster or firmer response.

The structure of a neural control system also abounds with nonlinear elements, usually "softened" switches which operate on a weighted sum of input variables values. Such switches are cascaded within "hidden layers" ­ though not many researchers have got beyond a single hidden layer. With such a structure, some elegant nonlinear controllers could be constructed in a very simple way ­ but most researchers seem to leap forward to a more appealing objective.

For many decades, the "Holy Grail" of control theory has been the "Universal Controller", a magic box which can be connected to any conceivable system and which will then proceed to learn how to achieve perfect control. The neural researchers see the neural network as the secret of that grail. The weighting coefficients become variables in their system, to be "trained" by nudging them in the direction which will reinforce any "correct" decision. The system performance will thus eventually be shaped by the reward function, rather than by analytic objectives.

It is easy to form the opinion that these researchers are playing "Let's pretend", as they speak of their systems as a fond mother would describe a precious child. "It learns to ...", "It does not have to know ...", "It explores ...". When tasks are simple, such as in devising a network which can determine the parity of five binary inputs, the suggestion that the data-space could be examined analytically to determine the distribution of local minima is met with amazement. The blindfold is an essential part of the game.

Many fuzzy-neural-genetic papers are characterised by little drawings of neurons, of double-helix genes or of "fuzzifier pyramids", accompanied by half a page or so of verbatim doctrine (or do I mean dogma?). Arguments are based on analogies with biological operations which are themselves shrouded in mystery, such as those of the cerebellum. It is easy to become cynical of them.

Sooner or later, though, you are bound to trip over the problem where fuzzy control comes into its own. I have a new research student who is basing his project on a bakery ­ a high-technology enterprise teeming with automation. While you can weigh a loaf to great precision, taste and texture are likely to be judged as "delicious", "rather ordinary", "chewy" and not as a neat 14.37. These inputs can be nothing but fuzzy.

The control system may well evolve with variable factors which will be perturbed by feeding back a reward for ''a good batch'' or a penalty for a batch which is unsaleable. Can this research be kept on the level of cold logic, or will it stray into the metaphysical?

Time will tell.