A rationale for a course of research into Nervous Nets
John A. deVries II -- April 21, 2000


To misquote the song "As Time Goes By", a car is but a car -- in other words, some technological elements of our society will probably never change size or shape to any radical degree. However, with the increasing feasibility of smaller and smaller devices eventually resulting in an all-pervasive nanotechnology, the way control is to be achieved must become smaller and simpler as well.

Some of the areas where Nv nets1have an essential edge over what has become the de facto use of microprocessors for control are simplicity, adaptability, and low power consumption. Simply to get a four-legged robot to walk with any facility requires a microprocessor consisting of anywhere from 10,000 to one million gates (circuit elements). The typical walker that uses a Nv net requires perhaps only 80 gates or so. Furthermore, the dynamic, analog implicit sensing of motor loads (called implex2) that occurs using Nv nets could only be emulated with a great deal of software on a microprocessor (if at all).

In nature, one finds many oscillators in the form of central pattern generators; the bicore is a human-made equivalent. Given matched components the frequency and duty cycle can be predictably set; substituting a component that reacts to external stimulus permits variability. Wilf Rigter has discussed both sorts of bicore on the BEAM emailing list3 and Wouter Brok's paper4 has also done so in some detail.

It is, for example, fairly easy to build a visual core or "head" using a bicore and a motor that follows a light or heat source and incidentally produces a signal representative of the direction. The paradigm of biological nervous systems then suggests a division of effort -- in this case, between sensor and motor functions. Most simply, one bicore can be used to process sensor information and another bicore can be used to provide motor control. If one accepts a hierarchy between sense and motion, the motor control bicore would act as a slave to the sensor processing bicore. Connecting two bicores so that one is master and the other is slave is also called embedding -- embedding one bicore into another in this manner is one possible starting point for building a kind of "spinal column"5. For an inductive model of "robotic evolution", this would represent the "3D" case.

One can imagine embedding (or surrounding) bicore within bicore within bicore and so forth that would give something that looks like a complicated bulls-eye. When considered in three dimensional space rather than on the plane of a piece of paper it could be seen as a perspective view of a tunnel, the centermost bicore being furthest away, containing the primary or first sensor. It can be compared to a simplified neural tube where each bicore is a ganglion which has motor neurons that lead out on each "side"; in other words, the basic model that controls worms and people alike. Implex can be used to alter the behavior of a single ganglion, providing the kind of control that is needed locally. Much of this is contained in "Living Machines"6 although the concept of the neural tube wasn't covered at that time. Presently, a robot named BEAMAnt 6.x is the most basic device using this two-bicore model. It has local "bump" sensors implemented using touch switches and Nu (integrating) neurons which alter the action of the motor control.

The next item, inductively speaking, is the "Snakebot" (also known as gPIM 2.0), created by Mark Tilden about 1995. Snakebot is a three-ganglion robot that doesn't have a head. As a result, it isn't merely a linear tube of three bicores but instead is a torus. The device is thus not much like a snake but it does a fair job of behaving like a blind worm. People have reported the sensation when it is held to be unpleasant. Perhaps snakes feel (muscularly) at least a little bit more pleasant than strong worms. gPIM 2.0 is still, however, merely a linear step. The longest example of this technology to date is the Lampbot 1.0 that is also known as the "Sidewinder". It consists of a head followed by a chain 8 segments long, each segment controlling one actuator and moves much like a lamprey

One can see that only the simplest and most linear of network topologies (including the single Nv neuron, loops, unterminated single chains, pairs of neurons and chains of paired neurons) have been explored so far. Unfortunately, very little of any other advanced research that might exist hasn't been published nor even described in abstracts. However, what information that does exist seems to predict a continued linear development of the concept.

The multitude of devices that might maintain the half-century dream of an automated house7 simply cannot be controlled using such a linear paradigm, both for the individual unit and for whatever "community" of units that would be required. Unless a new plan of analysis and development for nervous nets (or some other such scalable system) comes into being, we won't have the tools necessary to support such applications.

Possible research directions might include (but wouldn't be limited to):


1 US Patent 05325031, "Adaptive robotic nervous systems and control circuits therefore," Mark W. Tilden, June 28, 1994.


2 S. Still and M.W. Tilden: "Controller for a four legged walking machine," in: "Neuromorphic Systems: Engineering Silicon from Neurobiology", editors: L. S. Smith and A. Hamilton, World Scientific.


3 which, unfortunately, can only be found by searching the 13,733 articles on the alt-beam archive, January, 1999 to present day.


4 W. Brok, "The Suspended Bicore", July 30 1999, available in PDF form here.


5 An emailing list that attempted to cover this subject area can be found archived, June 1, 1999 and sporadically thereafter.


6 B. Haslacher and M.W. Tilden: "Living machines. Robotics and Autonomous Systems: The Biology and Technology of Intelligent Autonomous Agents", editor: L. Steels, Elsevier Publishers, 1995.


7 Ray Bradbury, "There Will Come Soft Rains", Collier's, May 6, 1950 (part of the "Martian Chronicles").