|
Mark W. Tilden,
Submitted for publication to the EANN '95 Conference Proceedings "Special Track on Robotics."
Nervous Net (Nv) technology is a non-linear analog control system that solves real time control problems normally quite difficult to handle with digital methods. Nervous nets are to Neural nets the same way peripheral spinal systems are to the brain. This work has concentrated on the development of Nv based robot mechanisms with electronic approximations of biologic autonomic and somatic systems. It has been demonstrated that these systems, when fed back onto themselves rather than through a computer-based pattern generator, can successfully mimic many of the attributes normally attributed to lower biological organisms. Using Nv nets, highly successful legged robot mechanisms have been demonstrated which can negotiate terrains of inordinate difficulty for wheeled or tracked machines. That non-linear systems can provide this degree of control is not so surprising as the part counts for successful Nv designs. A fully adept insect-walker, for example, can be fully controlled and operated with as little as 12 standard transistor elements.
Since the start of research in the winter of 1994, development of this technology has advanced to solving currently difficult sensory and cognitive problems. It is hoped in the coming years Nv systems may do for robot vision (amongst other disciplines) what has been done for autonomous robot vehicles, namely the reduction of currently complex systems down to an inexpensive but robust minimum. Further efforts are also being made to apply this control strategy to the expanding nanotechnology field. At the nanometer scale Nv's may prove more feasible than nano-computers for control of self-assembling micro structures. For now, however, Nv research concentrates on problems of scale invariance, proving by example (or exhaustion) this control system can work at all scales, types, and styles of robotic application.
The Nv control method could be adapted to most types of machine control, but it has been applied to autonomous robots because of the difficulty conventional control systems have solving the seemingly simple task of negotiating undefined complex environments. The 80 or so "biomorphic" robots (from the terms BIOlogy and MORPHology, and the Latin for "living" and "form") built so far are not "workers" in the traditional sense, but "survivors", in that they fight to solve the immediate problems of existence rather than procedural condition (i.e.: they do not follow the rules of an internal program that mimics the external world, but the world itself). Nv control architectures focus on adaptive survival rather than the performance of specific tasks. Once survivability is under control, goals can be superimposed and the machine used as a platform to carry sensors and conventional electronic intelligence. It is believed that these machines, although now in an early stage of development, can within a few years be brought to the point that they can serve as inexpensive, robust, and versatile carriers for a variety of instruments. A vast number of applications would then be possible, including the location and possible clearing of land mines from civilian areas, security, maintenance, medical and prosthetic applications (a cost-effective "walking wheelchair" for example), and even cars with on board "survival" instincts to save themselves, and their passengers, from damaging accidents. Though the Nv based legged devices built so far cannot go everywhere, they can certainly go places not currently accessible to wheeled or tracked vehicles of similar scale. It seems that for handling undefined environments, biomorphic designs are a very efficient and cost-effective approach.
Initially it was thought these devices avoided the problems of an internal world representation by using a reactive or behavior-based technique. Recent work has shown however that Nv biomorphs instead take a chaotic map of their surroundings onto their process control hierarchy (that is, they dynamically and efficiently adapt to the fractal complexity of their surroundings). This is due to the analog-electronic nature of the devices, the adaptive hardware of their structure, and the topological orientation of their interconnections. The defining characteristic of this adaptation is continuously updated by the immediate fractal complexity of the environment. These devices are "soft" designs, in that the environmental dimension must be absorbed, modified, and acted on for the devices to make successful headway through a complex world. These devices do not use "feedback" in the standard sense, but rather "implex", as the driving forces are augmented by perceived load rather than by a separate regulating path. The result is highly compliant, animal-like machine motions that "negotiate" rather than "bully" their way through environments, resulting in minimal damage to both world and robot.
We talk about these devices in the general sense because the precepts of their existence and subsequent design are based upon environmental macros, such as fluidity, turgidity, gravity, scale, materials strength, and many other factors. The power of biomorphic designs is that this information is used as the defining principles to shape appropriate survivor(s) for a particular environment. The machines that emerge are vastly different from any conventional robotic forms. We suspect, at least from the experimental evidence, that this technique embodies a new type of non-linear control paradigm, and at least an entirely new engineering discipline for the matching of competent machines to complex environments. Here, once the problems of existence are ratified, the devices can do unsupervised, long term work without human intervention (some devices have been in continuous operation for over 5 years).
The potential for this control paradigm is vast, but it is far from linear, and requires integrated design attempts to pull a competent ability from the Nv nets. To this end, the use of this technology to "evolve" machines from a lesser to a higher operational state has resulted in not only a wide spectrum of devices, but even completely different "species" of creatures, all evolved from a primal "genotype"; the single "cell" creature known as Turbot 1.0.
A further advantage is the speed at which this evolution has occurred, indicating that real-world Lamarckian evolution may match the success of many computer models yet seen. The diversity of this technique offers potential solutions to two main research fields, macro and micro robotics, and experimental work has been done to produce adept prototypes for both. The conclusions are that there may be some universal chaos-bounded concepts that bind survivor oriented designs, allowing for the creation and optimization of devices that can do work in any environment, under many situations, using chemically inert, and thereby relatively safe, control techniques (the idea of seeding a wheat field with pest-fighting silicon biomorphs to produce high yield, insecticide-free foodstuffs is an attractive example).
Considering that biomorphs may last long enough to replace most forms of long-term damaging chemicals (i.e.: pesticides, bleaches, medicines) the potential for the field really opens up. Deployed artificial chemicals perform a task in their immediate area of concentration, and then disperse into the environment where, after a time, they cannot be absorbed adequately. Biomorph machines, made from biodegradable silicon and trace elements, can be made gregarious so they do not of their on volition disperse, and can be absolutely controlled by conventional methods. Whether at micro or macro scales, these designs are not just capable, but competent. Furthermore, as they are self "programming" and non-reproductive, their behavior is both contained and predictable.
Nv biomorphs are something new with a demonstrable potential. Future work will concentrate on how this technology could fill in the cracks between science fiction and reality by finding out what is feasible now, and how to logically proceed to marketable, capable machines. In the coming years, it is hoped to be possible to demonstrate real machines to assess feasibility for macro and nano robotic applications. Expansions of the fields of robobiology, robomorphology, and artificial ecologys will be studied and published, along with extensions of this field from self-repairing processors, new computational paradigms, and even nanorobotic surgeons-in-a-capsule. Biomorphics is new, but it is slowly gaining the maturity and acceptance necessary to become a valid work tool.