Edited by
Robert A. Freitas, Jr. and William P. Gilbreath
Proceedings of the 1980 NASA/ASEE Summer Study
Held at the University of Santa Clara
June 23-August 29, 1980
NASA Conference Publication 2255
5.1 Introduction
As the cost of fossil-fuel energy continues to escalate and supplies of readily accessible high-grade ores and minerals gradually become depleted, the utilization of nonterrestrial sources of energy and materials and the development of a nonterrestrial industrial capacity become increasingly desirable. The Moon offers plentiful supplies of important minerals and has a number of advantages for manufacturing which make it an attractive candidate factory site compared to Earth. Given the expense and danger associated with the use of human workers in such a remote location, the production environment of a lunar manufacturing facility should be automated to the highest degree feasible. The facility ought also to be flexible, so that its product stream is easily modified by remote control and requires a minimum of human tending. However, sooner or later the factory must exhaust local mineral resources and fall into disrepair or become obsolete or unsuitable for changing human requirements. This will necessitate either replacement or overhaul, again requiring the presence of human construction workers with the associated high costs and physical hazards of such work.
The Replicating Systems Concepts Team proposes that this cycle of repeated construction may possibly be largely eliminated by designing the factory as an automated, multiproduct, remotely controlled, reprogrammable Lunar Manufacturing Facility (LMF) capable of constructing duplicates of itself which would themselves be capable of further replication. Successive new systems need not be exact copies of the original, but could, by remote design and control, be improved, reorganized, or enlarged so as to reflect changing human requirements. A few of the benefits of a replicative growing lunar manufacturing facility (discussed at greater length in secs. 5.4 and 5.5) include:
(I) The process of LMF design will lead to the development of highly sophisticated automated processing and assembly technologies. These could be used on Earth to further enhance human productivity and could lead to the emergence of novel forms of large-scale industrial organization and control.
(2) The self-replicating LMF can augment global industrial production without adding to the burden on Earth's limited energy and natural resources.
(3) An autonomous, growing LMF could, unaided, construct additional production machinery, thus increasing its own output capacity. By replicating, it enlarges these capabilities at an increasing rate since new production machinery as well as machines to make new machines can be constructed.
(4) The initial LMF may be viewed as the first step in a demonstration-development scenario leading to an indefinite process of automated exploration and utilization of nonterrestrial resources. (See fig. 5.1.) Replicating factories should be able to achieve a very general manufacturing capability including such products as space probes, planetary landers, and transportable "seed" factories for siting on the surfaces of other worlds. A major benefit of replicating systems is that they will permit extensive exploration and utilization of space without straining Earth's resources.
Figure 5.1. - Automated space exploration and industrialization using self-replicating systems.
5.1.1 Summary of Chapter Contents
The history of the concept of machine replication is reviewed in section 5.2. This theoretical background is largely a consideration of the work of John von Neumann—in particular, his kinematic and cellular models of automata self-reproduction. Post-von Neumann research is reviewed next, noting particularly the established theoretical capabilities of machines in the realm of general construction, inspection, and repair strategies. Such strategies may prove useful, even vital, to the successful design, realization, and operation of actual replicating systems.
Section 5.3 deals with the engineering feasibility of the concept of self-replicating systems (SRS). An attempt is made to confront two important general problems in creating a lunar replicating factory:
• Given that in theory, machines can construct duplicates of themselves, how might systems designers and engineers identify all functions which must be carried out to achieve machine replication and also develop the technological means by which to implement these functions?
• Given the constraints obtaining in the lunar environment, particularly in terms of the inventory of known kinds and quantities of naturally occurring raw materials and the existing repertoire of materials processing technologies, can all machine functions required both for production and for replication and growth be implemented?
To attack the first of these problems - identification of necessary functions for practical machine replication—the team proposes a specific phased demonstrationdevelopment scenario, described in section 5.3. For the second problem—establishing that machine replication can feasibly take place in the actual lunar environment - a strawman mission concept was employed. In this scenario, a 100-ton initial "seed'~ factory is planted on the Moon with access only to local resources and established materials processing techniques. The initial system should be able to successfully develop into an expanded machine system capable of conducting all functions necessary for autonomous replication, growth, and automated production and manufacturing.
The problem of "closure" is also considered at length in section 5.3. The issue of closure is whether autonomous manufacturing and construction systems can make available to themselves all of the materials, parts, and assembly techniques required for all internal operations. An iterative strategy is presented for detecting and eliminating closure gaps, and for optimizing the resulting augmented system.
Section 5.4 deals with possible applications of the SRS concept. Applications of replication technology include enormous gains in terrestrial industrial productivity (automation and computer-aided design and manufacturing), utilization of Solar System resources, orbital and planetary opportunities, and the possibility of interstellar exploration on a grand scale. Indefinitely large masses can be organized in extraterrestrial environments using self-replicating systems.
Section 5.5 deals with just a few of the many implications of SRS. The advantages of space-based replicathe manufacturing are considered, together with possible political, social, economic, cultural, and psychological consequences of the proposed SRS development program.
Section 5.6 sets forth in some detail how NASA can take action at once toward the achievement of the ultimate goal of establishing a replicating manufacturing facility. Suggested statements of work (SOWs) and a listing of institutions that might undertake the tasks outlined in the work statements are included. A series of specific conclusions and recommendations generated by the Replicating Systems Concepts Team are presented in section 5.7.
5.2 Theoretical Background
The notion of a machine reproducing itself has great intrinsic interest and invariably elicits a considerable range of responses—some directed toward proving the impossibility of the process, others claiming that it can be carried out, but almost all of them indicating an unwillingness to subject the question to a thorough examination. In discussing self-replication by automata it is essential to establish early rather important ground rules for the discussion. According to Kemeny (1955), "1f [by 'reproduction'] we mean the creation of an object like the original out of nothing, then no machine can reproduce - but neither can a human being....The characteristic feature of the reproduction of life is that the living organism can create a new organism like itself out of inert matter surrounding it."
Often it is asserted that only biological organisms can reproduce themselves. Thus, by definition, machines cannot carry out the process. On the other hand, others argue that all living organisms are machines and thus the proof of machine reproduction is the biosphere of Earth. Also, sometimes it is claimed that although machines can produce other machines, they can only produce machines less complex than themselves. This "necessary degeneracy" of the machine construction process implies that a machine can never make a machine as good as itself. An automated assembly line can make an automobile, it is said, but no number of automobiles will ever be able to construct an assembly line.
Another common argument is that for a machine to make a duplicate copy it must employ a description of itself. This description, being a part of the original machine, must itself be described and contained within the original machine, and so on, until it is apparent we are forced into an infinite regress. A variant of this is the contention that a machine not possessing such a description of itself would have to use itself for a description, thus must have the means to perceive itself to obtain the description. But then what about the part of the machine that does the perceiving? It cannot perceive itself, hence could never complete the inspection needed to acquire a complete description. (A simple counter is that the original machine might possess multiple perceiving organs, so that the perceiving could be shared.) Yet another related objection is that for the process to be carried out, the machine must come to "comprehend" itself - at which point it is said to be well known that "the part cannot possibly comprehend the whole." These disputations suggest that there is a very deep-seated resistance to the notion of machines reproducing themselves, as well as an admittedly strong fascination with the concept.
The Hungarian-American mathematician John von Neumann (1966), who first seriously came to grips with the problem of machine reproduction, once noted that it would be easy to make the whole problem go away. One could, for example, make the elementary parts of which the offspring machine was to be composed so complex as to render the-problem of replication trivial. In one example of this considered by the team, a robot required only to insert a fuse in another similar robot to make a duplicate of itself would find "reproduction" very simple (see sec. 5.2.3). As von Neumann also pointed out, it is equally useless to go to the other extreme and try to account for the placement of every atomic particle in the system—one would quickly become mired in incomprehensible detail. Even most lifeforms do not have DNA-encoded instructions for reproduction to this fantastic level of detail—their descriptions are largely at the molecular level.
As will be demonstrated presently, although reproduction may be transparently trivialized or intractably complexified, there appear to be no fundamental inconsistencies or insoluble paradoxes associated with the concept of self-replicating machines.
5.2.1 Von Neutnann's Contributions and Subsequent Research
John von Neumann began studying automata replication because he was interested in very complex machines and their behaviors. The early history of the theory of reproducing machines is basically the history of von Neumann's thinking on the matter, and this is reviewed below.
Von Neumann had a tremendous range of interests—he contributed to the logical foundations of quantum theory, was the co-inventor of the theory of games, and he worked on the Manhattan Project (contributing to the design of the implosion mechanism for the plutonium bomb). It is believed that his participation in the Manhattan Project and the tremendous volume of calculations necessary for bomb design led him into automatic computing. Hearing of the ENIAC computer project at the Moore School of Electrical Engineering at the University of Pennsylvania, von Neumann was fascinated by the potential of a computer very much faster than any of the devices that had previously been produced. In the early 1940s there existed only simple relay machines and analog devices such as the differential analyzer. But the new electronic machines that interested von Neumann promised to be perhaps millions of times faster than relay machines.
So von Neumann immersed himself in the ENIAC project, the first electronic computer program where some actual useful computing was produced. Late in 1945 and early 1946, the first problems that were put on ENIAC are believed to have been calculations involving the feasibility of a hydrogen bomb. Von Neumann, although he remained very much interested in nuclear energy and was appointed a member of the Atomic Energy Commission, became fascinated with the idea of large and complex computing machines. He devised the organization employed today in almost all general purpose computational machines—the so-called von Neumann concept of serial processing storedprogram or the "von Neumann machine." After that work was completed he began thinking seriously about the problems of extremely large machines—their reliability, programming, design, how to understand what they do— and he became involved with the many possible analogies to the complex behavior of living systems.
Von Neumann set for himself the goal of showing what the logical organization of a self-reproducing machine might be. He had in mind a full range of self-replicating machine models which he intended to explore, including the (a) kinematic machine, (b) cellular machine, (c) neurontype machine, (d) continuous machine, and (e) probabilistic machine. As it turned out, he ultimately was only able to produce a very informal description of the kinematic machine. Although he wrote a great deal on the cellular machine, his magt2um opus on the subject was left in the form of unfinished notes at the time of his death. Almost no work was done on the other three kinds of selfreproducing machines. For this reason, only the postulated workings of the kinematic and cellular machines are presented below, with brief comments on the other three types. For an additional review of these two models of reproduction, see Burks (1970).
In dealing with machines that could reproduce, von Neumann concluded that the following characteristics and capabilities should be demonstrable for each:
(1) Logical universality—the ability to function as a general-purpose computing machine able to simulate a universal Turing machine (Turing, 1936). This was necessary because SRS must be able to read instructions to carry out complex computations.
(2) Construction capability — to self-replicate, a machine must be capable of manipulating information, energy, and materials of the same sort of which it itself is composed.
(3) Constructional universality - in parallel to logical universality, constructional universality implies the ability to manufacture any of the finitely sized machines which can be formed from specific kinds of parts, given a finite number of different kinds of parts but an indefinitely large supply of parts of each kind.
(4) Self-reproduction —follows immediately from the above, since the universal constructor must be constructable from the set of manufacturable parts. If the original machine is made of these parts, and it is a constructable machine, and the universal constructor is given a description of itself, it ought to be able to make more copies of itself.
Von Neumann formally demonstrated that his cellular model of reproduction possessed these four properties.
Not much was done on a fifth property also believed to be important—evolution—though there have been some more recent results in this area. If one has a machine, and it makes a machine, which then itself makes a machine, is there any proof that the line of machines can become successively "better" in some fashion — for instance more efficient, or able to do more things? Could they evolve to higher and higher forms? This problem raises issues in learning, adaptation, and so forth, and was left largely untouched by von Neumann.
The kinematic machine. The kinetic machine is the one people hear about most often in connection with von Neumann's work on self-reproducing machines, probably because it received the earliest attention and publicity. John Kemeny (1955) produced a paper for the popular publication Scientifc American detailing this model, and a further description appeared in a paper by von Neumann (1 9s 1).
The notion of kinematic machine self-reproduction was dealt with by von Neumann only informally. The mathematician envisioned a machine residing in a "sea" of spare parts. The machine has a memory tape which instructs it to go through certain mechanical procedures. Using a manipulative appendage and the ability to move around in its environment, the device can assimilate and connect parts. The tape-program first instructs the machine to reach out and pick up a part, then to go through an identification routine to determine whether the part selected is or is not the specific one called for by the instruction tape. If not, the component is thrown back into the "sea" and another is withdrawn for similar testing, and so on, until the correct one is found. Having identified a required part the device searches in like manner for the next, then joins the two together in accordance with instructions.
The machine continues following the instructions to make something, without really understanding what it is doing. When it finishes it has produced a physical duplicate of itself. Still, the second machine does not yet have any instructions so the parent machine copies its own memory tape onto the blank of its offspring. The last instruction on the parent machine's tape is to activate the tape of its progeny .
Von Neumann's logical organization for a kinematic machine is not the only one possible, but probably is the simplest way to achieve machine self-replication. In its logic it is very close to the way living organisms seem to reproduce themselves (Dyson, 1979). One conceptual problem with the model is that the parts involved are supplied free to the machine, and those parts are of a relatively high order. The machine dwells in a universe which supplies precisely the sorts of things it needs as a kinematic device to make a duplicate of itself. This raises the issue of closure, a problem which is discussed and conceptually resolved in section 5.3.
The cellular model. Von Neumann evidently was dissatisfied with his original kinematic model because of its seemingly mathematical inelegance. This model of machine self-reproduction, while qualitatively sound, appeared not easily susceptible to mathematically rigorous treatment and so might not serve to convince a determined skeptic.
Stan Ulam, a Polish-American mathematician who had also worked on the Manhattan Project, suggested to von Neumann that the notion of a self-reproducing machine would be amenable to rigorous treatment if it could be described in a "cell space" format—a geometrical grid or tessellation, regular in all dimensions. Within each cell of this system resides a finite state automaton. These cell automata can only be affected by certain of their neighbors, and only in very specific ways. In the model von Neumann finally conceived, a checkerboard system is employed with an identical finite state automaton in each square (fig. 5.2). In this system, as it evolved with subsequent research, the cell-automata can be in one of 29 possible different states (fig. 5.3). Each automaton can communicate with its four cardinal direction neighbors. The state of a cell-automaton is determined by its own state and by the states of its cardinal direction neighbors.
Figure 5.2.- Finite state automation cellular space.
UNEXCITABLE ORDINARYTRANSMISSION
SPECIAL TRANSMISSIONCONFLUENT
SENSITIZED
Figure 5.3.- Twenty-r...~ von Neumann 's cellular automata.
At the beginning of operation, all but a finite number of the cell automata are in a "U" or "unexcitable" state. If a given cell is in the "U" state, and all its neighbors also are in the "U" state, then at the next moment of time, the given cell remains in the "U" state. Thus the "U" states can be viewed as representing undifferentiated, passive underlying substrate. Their passivity implies that they may in some cases serve as "insulation" surrounding more active cells in the system.
Then there are "ordinary transmission" cell states. These are states which direct their activity in each of the four cardinal directions. Each of these may be in an excited or quiescent mode, so there is a total of eight different kinds of ordinary transmission states. In addition, there are eight "special transmission states," similar to the ordinary states in that they also point in each of the cardinal directions and can be in excited or quiescent modes. The two basic kinds of transmission states—ordinary and special—differ in that the primary intended role of ordinary transmission states is the routing of informational signals, whereas the primary role of special states is to inject transforming signals into cell locations and thereby convert "U" cells into active elements (or, if need be, convert active elements back into "U" cells).
The system also has four "confluent" states. They are activated if they receive signals from all cells in their neighborhood which are directed toward them. If activation occurs, then after two moments of time they emit signals outward toward any cell in their neighborhood which does not have a transmission directed toward it. Thus, confluent cells can serve as "and" gates, and as wire branching elements. Since they do not emit their output until two moments of time have elapsed, the confluent cells can also be employed to create time delays in the transmission of signals. The eight remaining cell states of the 29 originally employed by von Neumann are of less importance. These are temporary cell states which arise only as the operational states are being created from "U" cells.
Von Neumann first showed how to design a general purpose computing machine in his cell space system. He did this by showing the design of various basic organs — "pulsers" to emit any desired finite train of pulses upon activation, "periodic pulsers" to emit repeated trains of desired pulses after activation until signaled to stop, "decoders" to detect the presence of certain patterns of pulses, and the like. Using these organs, von Neumann developed a design for the control portion of a computing machine in one region of the cell space. He then showed how to organize an adjacent but indefinitely extendable portion of the cell space into a memory or information storage unit, which could be accessed by the control unit.
For the process of construction, von Neumann designed a construction unit, which, taking instructions from the memory unit, could send out a constructing arm (by cteating an active pathway of transmission cells into a region of "U" cells) and at the end of the arm, convert "U" cells to the cell types specified in memory (see fig. 5.4). He showed that this constructor could create any pattern of passive cells whatsoever. Thus, he had designed with mathematical rigor a universal constructor, relative to all possible passive configurations of cells in the cell space.
Since the parent machine itself can be created in passive form, it can make a duplicate of itself by the following process. The parent machine is supplied initially with instructions to make a duplicate of its control, construction and memory units (the memory unit initially is empty). After it completes this major construction phase, the instructions call for the parent machine to make a copy of the instructions in its memory and to feed into the memory unit of the newly constructed machine. Then the parent machine activates the heretofore passive offspring machine, and withdraws the constructing arm. At that moment the offspring is a duplicate, in all respects, of the parent at the time the original machine commenced its reproducthe activities.
Figure 5.4.- Universal construction in the cellular model of machine self-reproduction.
Critique of the cellular model. Although the 29-state von Neumann cellular array system permits a more elegant mathematical approach to the problem of machine construction and self-reproduction, it is more difficult to envision an actual useful physical implementation of the process (compared, say, to the kinematic model of replication). The entire cell space enterprise proceeds in a highly constrained artificial environment, one which is very special despite some features relating in a general way to the natural world. For example, the movement of objects in space, a ubiquitous and familiar phenomenon in the real world, becomes a complex process of deletion of cell states at one location and re-creation of these states at some other location.
There is also an assumption of synchronous behavior throughout the system. All cells, no matter how distant, are subject to change of state at the same instant, a property which would be difficult to implement in any practical large cell space. Indeed, the requirement of a source of clocking pulses violates the array symmetry which makes the cell space notion an attractive object for mathematical treatment.
It is also very difficult to design machines of interest which can be embedded in the cell array format. To make design and embedding easier, a higher-level machine design language would have to be created. It is likely that, rather than undertake that task, one would first redesign the underlying cell space properties to rid the system of the deficiencies already noted.
For instance, one might wish to introduce a new primitive cell state in the system to permit signals to cross without interference. A "wire-crossing" organ can be devised using only the original von Neumann primitive cell types, but this introduces an unnecessary complexity into the machine design process since the organ contains initially active cell states whose creation involves considerable extra care to avoid the propagation of spurious signals. This extra care is especially critical because the cell system, as von Neumann originally constituted it, is highly susceptible to signal errors. (He undoubtedly intended his probabilistic machine model to mitigate this sensitivity and fragility.)
The cell space system has very limited capacity to detect the states of cells. It has some capacity to detect states, for this is required in the operation of the memory unit. But a machine cannot analyze an arbitrary encountered cell to determine what state it is in, thus cannot "read" the states of an encountered machine. This inability severely restricts the capacity of cell-space machines to repair other machines or to attempt self-repair. Such limitations also are evident in the construction process, where the constructing machine must assume that the region in which a new machine is to be created consists entirely of elementary quiescent cells. Should this not be the case, there is no systematic and complete way to detect it. A machine can send destruction signals into cells to reduce them to the quiescent form. Unfortunately, in some cases one must know the state of the cell ahead of time in order to determine what destructive signal must be sent to destroy it.
Finally, all machines that can be produced in von Neumann's cell space system are essentially information transactional devices. Even construction is, in this context, a form of information processing. Physical construction and material transformations can possibly be viewed as informational processes but, in a practical sense, the cell-space notion is far from providing a readily useful paradigm of actual manipulation and transformation of physical materials.
Von Neumann's other self-reproducing machine concepts. In addition to his kinematic and cellular models, von Neumann planned to examine three other models of self-reproducing machines. These were to be a neuronal or "excitation-threshold-fatigue" model, a continuous model, and a probabilistic model. Von Neumann is not known to have left any completed work whatsoever on these models at the time of his death, so his intentions are almost entirely a matter of conjecture.
Following Burks' speculations on this matter (von Neumann, 1966), we can guess that von Neumann's neuronal system might have been a version of the cell-space model in which the individual cell automata in the space were to be constructed of neuron-like elements. This would have been a rather straightforward process, as it is well known that idealized neurons of the McCulloch-Pitts (1943) variety can be employed to implement the kinds of logical gatings and delays called for in the 29-state cell automaton system. The reason for employing neuron-like elements seems mainly an attempt to recast the model in a more "biological" vocabulary.
Von Neumann's postulated continuous model might have been an attempt to comprehend machine reproduction in an even more biological format. The usual mathematical tools for handling actual neuron activity are differential equations expressing the electrochemical flows through and along neuron soma and axons. Thus the actions of cell automata (implemented with neurons) could be expressed by sets of differential equations. In this way the more highly developed tools of mathematical analysis might be employed in representing the behavior of the machine system, in contrast to the use of combinatorics which von Neumann himself characterized as one of the most intractable of mathematical specialties.
Finally, in his proposed probabilistic model von Neumann perhaps intended to consider using whole congeries of neuron-like elements in implementing the behaviors of what in the neuronal model could be carried out by single neurons. By employing redundancy techniques similar to those described in his classic paper on reliability, von Neumann (1956) may finally have hoped to design a reliable, biologically oriented, self-reproducing machine characterizable by differential equations. We can only guess.
Alternative cell array systems. Worlc on cellffpace automata systems in the period following von Neumann's contributions has taken several research directions. The underlying cell-space notion of a homogeneous medium with a local transition function that detemlines global properties has been employed in numerous modeling and simulation projects. For example, weather simulations use the idea of connected cells, the changes of each cell state described by a set of differential equations. Studies of the flow of excitation in heart tissue, the dispersal of medicinal drugs, and pattern recognition all have employed the cellspace concept. Cell spaces also have been investigated as abstract mathematical objects where, for instance, one tries to determine whether from every mathematical pattern all other pattems can be attained, and whether there are some patterns not attainable at all by means of the transition function, and various other specialized questions.
Some work in cellular automata has attempted to carry forth the von Neumann program of machine construction and self-reproduction. For instance, Codd (1968) recapitulated the von Neumann results in a simpler cell space requiring only 8 states rather than 29. This produced a machine design recognizably closer to that of present-day computing machines. Myhill (1970), trying to mitigate the artificiality of the indefinitely extended pre-existing cell space, designed a system in which componentry was drawn into a cell-grid system and was then employed as machine constituents somewhat as biological cell constituents might be drawn through a membrane to be used at an intracellular work site. Arbib (1966), attempting to make the movement of cell machines a less cumbersome matter, designed a cell-space system in which cells and blocks of cells might be joined together by a "welding" operation, thus becoming "co-moving" configurations.
Smith (1970) and Banks (1970) introduced additional simplifications to the cell-space notion, showing that the von Neumann program could be recapitulated in underlying cell spaces of an extremely elementary sort. Indeed, the so-called "Game of Life" designed by Conway (Gardner, 1971) is a cell-space system which, despite its very simple transition rules, has been claimed to be capable of expressing both universal computation and construction. (The game involves a checkerboard cell array with cells in one of two states, "0" or "1." A point whose state is "0" will change to state "1" if exactly three of its eight neighbors are in state "1." A point whose state is "1" will remain in that state if two or three of its neighbors are also in state "1." In all other cases, the state becomes or remains "O.")
Later research on self-reproducing automata. By the late 1960s, the original von Neumann program of machine construction and reproduction had been largly abandoned, although investigation of cell-space systems as abstract mathematical entities or as vehicles for "spatial" modeling and simulation has persisted. Indeed, research in the latter field has been especially vigorous and prolific - one recent author lists over 100 references for cell-space imaging applications (Preston et al.,1979).
Von Neumann's kinematic machine construction system appears to have had no intellectual progeny whatsoever. This is somewhat misleading, since practical application of computers to manufacturing and the persistent human interest in and investigation of robot mechanisms have, without explicit connection to von Neumann's earlier work, prepared the ground for a possible implementation of a hybrid computer/kinematic model of machine construction and reproduction.
The theoretical work of this later period, explicitly derived from von Neumann's research effort, has focused mainly on the molecular biological analogies that can be drawn. For example, in a series of papers Laing (1975, 1976, 1977, 1978, 1979) employs a hybrid cellularkinematic model of machine construction and shows that neither existing natural nor artificial machines need be bound to follow the "classical" reproductive paradigm. In the classical paradigm, a program (DNA in living systems) is first interpreted to construct a machine (protein synthesis in lifeforms) and then is read a second time to make a copy of the program for insertion into the newly constructed duplicate machine (DNA replication in living cells). The principal contribution of Laing is to suggest reproductive strategies other than direct analogues to the known biological process. In this new conception, a machine is able to identify all of the components of which machine systems consist (not merely a subset as in the von Neumann cell system) and can access all of an existing machine structure without requiring dismantling of the system (as would be required in the von Neumann model).
Once this and other similar advanced concepts are brought to bear on the problems of machine reproduction, many alternative reproduction strategies become immediately apparent. A selected few of these are reviewed in the following section.
5.2.2 Alternative Replication Strategies
A number of alternative automata reproduction strategies have been suggested in the decades following the completion of von Neumann's work. Major strides have been made in the scientific understanding of the processes of biological reproduction at the molecular or biochemical level. Recent research has demonstrated the theoretical possibility of inferring structure and achieving selfreplication without first possessing a complete selfdescription. This suggests an enormous range of new machine capabilities which possibly may be technologically exploited in the future, according to specific rules and multiplication strategies for optimal deployment.
Biological reproduction. Biological reproduction is thought to obey the following underlying logical paradigm. The basic genetic program (encoded in the genetic DNA) is employed to make a copy of the same information in a slightly different medium (RNA). This modified form of the genetic program is transported to a work site within the cell where, with the aid of cellular enzymatic machinery, the RNA is interpreted as coding for amino acid strings (proteins). The protein produced plays two major roles: (1) it constitutes the basic structural material of living organisms, and (2) certain smaller and variably active proteins (enzymes) control the metabolic, interpretive, and constructive actions of the system.
When the genetic code embodied in the RNA has been read and acted upon, the machinery construction phase is complete. The cell must then undertake the copying of original genetic material (the DNA) to provide offspring organisms with the necessary instructions. This copying process is the well-known DNA replication phase, in which DNA (in most cases a twisted pair of complementary DNA molecules) untwists to permit new nucleotides to match with existing separated strands to form two twisted pairs of DNA. Reproduction is completed when the newly produced and original organism machineries are divided up, one DNA program remaining with each.
This highly simplified description of biological reproduction is offered only to illustrate the underlying logical strategies: (I) follow instructions to make machinery, (2) copy the instructions, (3) divide the machinery, providing a sufficient set in each half, (4) assign a set of instructions to each half, and (5) complete the physical separatiOn.
Von Neumann 's automata reproduction. Von Neumann's automata reproductive process closely mirrors the biological one. In the original model, instructions exist in two copies. One of the copies is read and acted upon to construct another machine, sans instructions. The second copy is then read and copied twice, and this double copy is inserted into the passive constructed offspring machine which is then turned on and released, thus completing the act of reproduction.
There is no logical necessity for having two sets of identical instructions. Von Neumann employed two copies of the instructions because it eliminated the criticism that the instructions might, in the first (construction) phase, become corrupted and so not be able to transmit a true version for the use of offspring machine. Also von Neumann feared that there might seem to be a paradox in the program acting upon itself to make a copy of itself. There are, however, ways by which a program can successfully be made to make a copy of itself, and indeed many such programs, though exceedingly simple, have already been written (Burger, Brill, and Machi, 1980; Hay, 1980). Another solution is to provide the machine proper with an automatic "wired-in" copy routine which the program calls for at the proper time.
Simplifed von Neumann automata reproduction. Consider a single instruction tape, and a constructor machine which reads the instructions once to build the offspring machine and again to make a copy of the instructions for the offspring machine. Notice that although the instructions available to the system yield a duplicate of the original system, this need not be the case. Thus, in the biological example, even though some DNA made available to a cell does not code the instructions for a duplicate cell, the cellmachine still may proceed to obey the instructions. This means that a cell can generate offspring not only different from itself and its normal constituents and products, but even inimical to it. This is precisely what happens when a virus possessing no metabolic machinery and no enzymatic protein machinery to read DNA or to manufacture anything parasitically insinuates itself into a host cell. The virus co-opts the host cell's interpreting and manufacturing capacity, causing it to make virus particles until the cell fills with them, bursts open, and is destroyed. The greatly multiplied viral agents are then free to parasitize other cells.
In artificial systems as well, machines may read and interpret instructions without knowing what they are being called upon to do. The instructions might call for some computational, constructional, or program-copying activities. The machine can make machines unlike itself, and can give these "unnatural" offspring copies of the instructions which were employed in their manufacture. If the offspring are also equipped to read and follow instructions, and if they have a constructional capability, their offspring in turn would be replicas of themselves—which might not resemble their "grandparent" machine at all. Thus, an original construction machine can follow instructions to make an indefinitely large number of diverse machines, that are like or unlike themselves, capable or not capable of constructing, reproducing, etc. And though a universal constructing machine might make large numbers of "sterile" machines, if it should once make a duplicate of itself which is also equipped with the instructional program for making duplicates of itself, the process can become "explosive." Such machines would tend to drive out all other "species" not possessing this reproductive "autocatalytic" property. Thatcher's variant: inferring structure. Thatcher (1970) showed that a machine need not have an explicit construction program made available to it initially in order to create a duplicate of itself. First, it is sufficient that a machine can secure a description of itself (in place of instructions) if the machine is equipped with the capacity to read the description and convert this into the necessary constructive actions. Second, using a result obtained by Lee (1963) and himself (Thatcher, 1963), Thatcher showed that such a machine need not have its description loaded beforehand into its accessible memory organ. Instead, the machine has a partial self-description hard-wired into itself in the form of circuits which, when stimulated, make the description available to the machine in its accessible memory organ. These data describe all of the machine except the hardwired part which was stimulated to emit the description in the first place. The problem then, for the machine, is to obtain the description of this hidden part of itself. Lee and Thatcher showed that this section of the device can be constructed in such a simple fashion that the system can infer how this part must have been constructed merely by examining the consequences of its actions (e.g., the partial description it produced). After inferring the nature of this hidden part of itself, the machine possesses a complete selfdescription and can then follow von Neumann's paradigm for reproduction.
The principal practical significance of this form of automata replication is that it reminds the designer that the information required for machine construction (whether reproduction or not) need not be in the form of instructions for constructions but can be in the form of a description. Moreover, the description need not even reside in an accessible organ such as memory registers but may be embedded in "inaccessible" hardware. The hypothetical infinite regress likewise is shown to be baseless—it is possible for a machine to have within itself only a part of its own description, and from this to infer the rest.
Reproduction by component analysis. In von Neumann's cellular system, an embedded machine cannot send out an inspection arm to an encountered machine to identify all of its states. However, the cell-space system could be redesigned to permit this. In such a system an analyzing machine could examine an encountered passive machine and identify the type and location of all its cell-automata. (The analyzer might of course have to penetrate the machine, thus altering its automaton states, so the inspecting arm would have to send out appropriate restoration construction signals.)
In von Neumann's kinematic model a machine ostensibly could identify all parts of the system and thus determine the type and location of all components. This opens the possibility that a machine system might, for example, reproduce essentially two machines — one active, the other passive or able to assume passivity under a signal from the active machine. This possibility and others have been explored by Laing (1975, 1976, 1977, 1978, 1979) in a series of papers presenting alternative reproductive strategies which include the following:
• Beginning with two identical machines, one active and one passive, the active machine "reads" the passive machine twice, producing one active and one passive machine, thus completing reproduction.
• Beginning with two machines (not necessarily identical) one machine reads the second, and makes a duplicate of it. Then the second reads the first, and makes a duplicate of it, active and passive status being exchanged.
• By combining the capacity of machines to read machines with the Thatcher result, one can hardwire a machine to construct a second machine which is a duplicate of the original except for the hardwired part which produced the second machine. The original machine then "reads" the newly constructed partial duplicate, and infers what the missing hardwired part must be. The original machine then constructs the missing part, completing the reproductive process. This result explicitly confronts and overcomes the "necessary machine degeneracy" criticism of automata self-replication.
Machine reproduction without description. In the machine reproduction schemes explained thus far, some arbitrary part of the machine which cannot be inferred is always made explicitly available in memory initially, or is implicitly made available iwn memory or for inspection by means of an internal wired-in memory, also not directly accessible. Laing (1976) showed that even this wired-in description is not necessary. In effect, a machine can carry out a self-inspection which can yield a description which in turn can be made available to the machine in constructing a duplicate of itself.
The process begins with a wired-in construction routine which produces a semiautonomous analyzer machine. This analyzer moves over the original machine and identifies the type and location of its componentry. This is reported back to the original machine, which uses this information to make a duplicate of itself. Thus, though it may be that a part of a machine "may not comprehend the whole" in a single cognitive act, a part of a machine can examine in serial fashion the whole machine, and in time can make this information available to the machine for purposes of replication.
Exploitation of basic machine capabilities. The "simplified von Neumann" automata reproductive strategy — whereby a machine employs a stored program of instructions to make other machines (including duplicates of itself) and then also provides the program or parts of programs of instructions to newly constructed machines— should probably be the central strategy for any actual physical machine reproducing systems. -The other strategies are, from most points of view, more complex than this and thus perhaps are less preferable. The virtue of the alternative strategies is not as practical ways of implementing machine reproduction but rather in suggesting many basic capabilities, which, in a complex automated replicating LMF, may be usefully employed. The following are some of the behaviors of which, under suitable conditions and design, machines are actually and potentially capable:
(1) A machine can be "hard-wired" to carry out a computation.
(2) A machine can be programmed to carry out a computation.
(3) A machine can be a general-purpose computer, in that it can be given a set of instructions which will enable it to carry out the computation of any other computer. Alternatively, a general-purpose computing machine can be given the description of any other computing machine, and can carry out the computational actions of the machine described .
(4) A machine can be hard-wired to carry out a construction activity.
(5) A machine can be programmed to carry out a constructional activity.
(6) A sufficiently complex machine can be a generalpurpose constructor, vis-a-vis a set of machines, in that it can be given a set of instructions which enables it to carry out the construction of any of the set of machines. Alternatively, a machine can be given the description of any machine of the set, and can, from this description, construct the machine described.
(7) A machine can construct a duplicate of itself, including the instructions or description used to guide the construction process.
(8) A machine, given a coded set of instructions for machine actions, or a coded description of a machine, can make a copy of the instructions or coded description.
(9) A machine, given a coded set of instructions for machine actions, can infer the structure of a machine which can carry out the actions described, and can construct such a machine.
(10) A machine, given a coded set of instructions for a machine, or a description of a machine, can carry out the actions of the machine whose instructions are given or whose description is supplied.
(11) A machine, given the instructions for or the description of an unknown machine, can examine the instructions or description and can (a) infer some of the properties of the machine, (b) simulate the actions of the machine, (c) construct the machine, and (d) observe the actions of the constructed machine.
(12) A machine can determine the component types of encountered machines.
(13) A machine can determine the structure (the component type and arrangement of components) of encountered machines.
(14) A machine can thus obtain a structural description of an encountered machine and simulate its actions, construct a duplicate, and then observe the duplicate in action.
(15) A machine can possess a copy of its own description, perhaps stored in a memory organ.
(16) A machine can obtain a copy of its own present structure. Note that the present structure of a machine may deviate from the original design, and also from its present stored description of itself (which may be out of date).
(17) A machine can compare the stored description of itself with the description obtained by inspection, and note the discrepancies.
(18) A machine can make a duplicate of itself on the basis of its stored "genetic" description or on the basis of its present (possibly altered) structure. This latter is an example of transmission of acquired characteristics.
(19) A machine can examine duplicates of itself constructed on the basis of an examination of itself, and note the discrepancies.
(20) The duplicates made from either of these two bases (genetic and observed) can be set in action and observed.
(21) For diagnostic purposes, the two kinds of descriptions can be compared, the two passive structures compared, the two kinds of structures in action observed and compared. The basis for machine self-diagnosis is thus available.
(22) A machine noting the discrepancies between two machine descriptions, or machine structures, or two machine behaviors, can in some cases act so as to resolve the discrepancies. That is, a machine in some cases can repair or reject or reconstruct deviant machines (including itself) .
(23) A machine encountering an "unknown" machine can observe the behavior of that machine and compare this to the behavior of other machines, both directly and by simulating the behavior of those machines for which it already has or can obtain descriptions.
(24) A machine encountering an unknown machine can examine the structure of the machine and obtain a structural description which can be compared with other structural descriptions.
(25) Encountering an unknown device, a machine can use the structural description of the unknown to simulate its actions. These simulated actions can be compared to those of other machines whose descriptions are stored or which can be made available.
(26) Having the description of an encountered device, a machine.can construct a duplicate of it. This duplicate can be set in action and observed, and its behavior compared with the behavior (actual or simulated) of other machines.
(27) The structure and behavior of encountered machines can be compared with those of known useful or benign machines, including that of the inspecting machine itself. This comparison, and the degrees of similarity discerned, can serve as the basis for a subsequent policy of "friendship," "tolerance," "avoidance," "enmity," etc.
(28) The descriptions of encountered machines can be incorporated into the reproductive construction cycle so that these new machines or their features become part of the continuing and evolving machine system. This is an analogue to biological symbiosis.
Machine multiplication strategies. In describing the logical process of machine reproduction we have concentrated on the means by which the parent system could come to possess the information needed to carry out a replication and the associated question of how offspring would if necessary acquire the programs needed to continue the machine reproduction process. Although these questions, logically, are at the heart of machine replication, they leave open many issues concerning creation and siting of new machine systems as well as the ultimate fate of such systems.
This matter can be approached by considering certain biological analogues to the machine situation. In the known biological realm, all living organisms use the same underlying reproductive logic of protein synthesis and nucleotide sequence copying but employ vastly different broad strategies in producing more of their own kind.
One strategy is seen in the case of seed-bearing plants (as well as most fish and insects), in which vast numbers of "minimal" genetic packets are produced by the parent system and dispersed in the hope that a sufficient number will, largely by chance, find an appropriate site at which to survive and complete growth and development to maturity. At the other end of the scale is human behavior, whereby "construction" and nurture of the offspring may continue under the control and protection of the parent system until near maturity.
The particular multiplication strategy for artificial reproducing systems must of course be adjusted to intentions. The swift utilization of large rich environments might justify a "seed" dispersal strategy, with early maturity of new systems so as to retain a high reproductive rate. On the other hand, an environment consisting of scattered pockets of valuable resources, or situations with less pressure for immediate "explosive" utilization might suggest fewer offspring, possibly more fully developed in regard to their capacity for seeking out and efficiently utilizing the scarce resources available. In this case, the offspring might also be expected to receive longer tutelage from the parent system or from outside controllers (such as humans).
Similarly, the presence of a large contiguous valuable ore body might dictate the extensive ramification of a single machine factory system consisting of many laboring submachines. The model of a colonial organism such as coral, or of a social insect such as ants or termites, might make more sense. Zoological and sociobiological studies of animal and plant multiplication strategies may prove valuable in suggesting optimal machine system growth and reproduction strategies. One important difference must be borne in mind: biological organisms often have adapted their strategies to compete with other organisms, as well as to survive in a world where resources are renewed at certain rates over varying seasons. Some of these factors may be nonexistent or present in very different form in a nonterrestrial machine-inhabited environment.
A few questions that should be considered in determin ing optimal replicating machine behavior include:
• How large should a system be allowed to grow?
• How large should a system grow before it reproduces.
• What sorts of offspring (e.g., minimal vs mature should be produced? A mixture?
• How many offspring should be produced? How manz offspring should be produced from a single parent machine?
• When should offspring be produced?
• Where and how should offspring be sited? Specific sites? Near? Far? Randomly dispersed?
• What offspring transport mechanisms should b/ employed? Should new systems be mobile? Under own control? Parent? Human operator?
• When should sited machine systems be turned off? Abandoned? Should lifespan of a machine system be a function of time alone? Reproductive life? Exhaustion of local resources? Work experience and use? Detection of malfunction? When should subsystems be turned off? What growth and death patterns of individual machine systems should be adopted?
• What should be done with unsited offspring systems? Allowed to wander indefinitely?
• What should be done with outmoded machine systems? Dismantle them? Abandon them?
Intergeneration information transmission among replicating machines. Throughout most of the present discussion it has been assumed that the goal was to have the parent machine transmit to its offspring machine the same genetic information it received from its parent, regardless of the logical strategy of reproduction employed. This genetic fidelity is not necessary or even desirable in all cases. Nor mally the parent should transmit all information necessary for offspring to do their jobs and to construct further offspring in turn, but beyond this simple requirement there are many alternatives. For example, a parent machine might augment its program during its lifetime with some valuable information, and this augmented part of the program could then be transmitted to its offspring.
A few possible variations of interest include:
(1) The parent machine program is not altered in the course of its lifetime and is transmitted unaltered to offspring.
(2) The parent machine program is altered (e.g., by intervention, or by some machine adaptive process of a more or less complex sort) during the course of its lifetime, but again only the program originally received from the parent is transmitted to the offspring.
(3) The parent machine program is altered during the course of its lifetime, and the altered program is transmitted to the offspring machine. The parent machine (being a constructing machine) may make changes in its structure beyond those called for in its received genetic program.
(4) Changes in parent structure are not made part of the offspring structure.
(S) Changes in parent structure are made part of the offspring structure.
(6) Changes in parental structure are not made part of the offspring structure, but are made part of the offspring genetic program. Thus, the offspring can, under its own control, modify its structure to conform to that of its parent machine.
5.2.3 Information and Complexity in Self-Replicating Systems
The design and implementation of a self-replicating lunar factory represents an extremely sophisticated undertaking of the highest order. It is useful to consider the complexity of this enterprise in comparison with the information requirements of other large systems, natural or artificial, replicating or not (Stakem, 1979).
It is not irnmediately clear what the proper measure should be. One way to look at the problem of machines reproducing themselves is to consider the flow of information that occurs during reproduction. A fully generalized self-replicating system could possess a reproductive behavior of such complexity that the information necessary to describe that behavior is complete to atomic level specifications of machine structure. Such a machine has behavior so complex and complete that it might produce a copy of itself almost from complete chaos—say, a plasma containing equal concentrations of all isotopes. In this case the machine reproduction is essentially complete—given sufficient energy, the system can make copies of itself in any arbitrary environment even if that environment contains virtually no information relevant to replication.
At the other extreme, consider a long row of Unimate PUMA-like industrial robots side by side, each requiring merely the insertion of a single fuse to render it functional. The first working robot, its fuse already in place, seeks to "reproduce" itself from a "substrate" of dormant machines. It accomplishes this by reaching onto a nearby conveyor belt, picking up a passing fuse part, and plugging it into the neighboring robot. The adjacent machine now begins to function normally as the first (indeed, as an exact duplicate), so it can be said that in some sense the first machine has reproduced itself. Before the reproductive act there was no second working robot; afterwards, one exists. However, this is almost the most trivial case of replication imaginable, since the substrate for reproductive activity in this case completed machines lacking only fuses—is extremely highly organized. Hence, the operative complexity resides in the substrate, and the action of the machine in "making a new machine" is trivial.
This latter example may be compared to the case of a bacteriophage. The phage particle infects a healthy bacterium, using the captive cellular machinery to manufacture new virus particles. Only the DNA of the virus enters the bacterium, instructing the cellular machinery to make new viral DNA and to interpret the DNA to create protein and polysaccharide components which form the coat or carrier of the viral DNA. Thus the foreign DNA, like the PUMA robot which inserts fuses to "self-replicate," must situate itself in a very rich complex environment, one already containing a great deal of machinery and information. In this case, the complexity of the virus-making enterprise probably can be gauged by the length of the viral DNA inserted into the host cell, just as the true complexity of the fuseinsertion behavior can be gauged by the length of the program needed to permit location of the supply of fuses and the fuse holder on an adjacent machine in physical space, and to insert the part properly. It is suggested, therefore, that the length of the shortest program which can carry out the process of replication may be an appropriate measure of the complexity of the task.
For instance, in the case of the von Neumann cellular reproducing system each part is already located in its proper place in space, but signals must be injected into that space to cause it to take on the properties desired in the offspring machine. It has been estimated that such a reproducing machine might consist of a minimum of 105 cells, with offspring cell type and location the principal parameters which must be specified for each. The length of the shortest program would represent perhaps 106 bits of information (Kemeny, 1955).
If the construction of a replicating growing lunar factory was purely a matter of machine parts assembly, then the length of the replication program could be determined by the necessity to locate various required parts in the environment and then to specify and execute the proper placement of each part to construct the desired system (Heiserman, 1976). However, it is likely the reproductive process will be vastly more complicated than this, since it is not likely that all parts can be supplied "free" from Earth. If the lunar factory must begin, not with completed machines or parts, but rather with a raw lunar soil substrate, the task quickly becomes many orders more difficult—though not impossible. Based on the estimates outlined in section 5.3 and the appendixes, the lunar factory replication program length should not exceed roughly 1012 bits of information. This compares to about 101° bits coded in the human genome and about 1014 bits stored in the human brain. Terabit (1012 bits) memories are considered state-of-the-art today.
Complexity of a selt-replication program may also be viewed as an index of versatility or system survivability. The more complex the program, the more likely it is that the machine system can bring about its own replication from increasingly disordered substrates. This is an interesting observation because it suggests that reproduction is an activity defined along a broad continuum of complexity rather than as a single well-defined event. Both the chaosreplicator and the fuse-insertion robots described above perform acts of self-reproduction. Fundamentally, these systems differ only in the degree to which they are capable of bringing order to the substrate in which they are embedded.
It is interesting to note that human beings fall somewhere in the middle of this broad reproductive spectrum. A 100 kg body mass, if composed of purely random assortments of the 92 natural elements, would contain roughly 1027 atoms and hence require about 1028 bits to describe. Yet a 100 kg human body is described by a chromosome set containing just 101 ° bits. The difference must be made up by the "substrate" in which people are embedded - a highly ordered rich environment, namely, the Earth. Human beings thus are conceptually remarkably similar to von Neumann's kinematic self-reproducing automata, moving around in a "stockroom" searching for "parts."
5.2.4 Conclusions
The Replicating Systems Concepts Team reached the following conclusions concerning the theory of machine reproduction:
(1) John von Neumann and a large number of other researchers in theoretical computer science following him have shown that there are numerous alternative strategies by which a machine system can duplicate itself.
(2) There is a large repertoire of theoretical computer science results showing how machine systems may simulate machine systems (including themselves), construct machine systems (including machine systems similar to or identical with themselves), inspect machine systems (including themselves), and repair machine systems (including, to some extent, themselves). This repertoire of possible capabilities may be useful in the design and construction of replicating machines or factories in space.
5.3 Feasibility
The design and construction of a fully self-replicating factory system will be a tremendously complicated and difficult task. It may also be fairly expensive in the near-term. Before embarking upon such an ambitious undertaking it must first be shown that machine self-replication and growth is a fundamentally feasible goal.
5.3.1 Concept Credibility
The plausibility of the theoretical notion of selfreplicating machines already has been reviewed at length (see sec. 5.2). It remains only to demonstrate concept credibility in an engineering sense (Bradley, 1980, unpublished memorandum, and see appendix SA; Cliff, 1981; Freitas, 1980a; von Tiesenhausen and Darbro, 1980) - that is, is it credible to consider building real physical machines able to replicate themselves?
The credibility of any design proposed for such a machine or machine system depends first and foremost upon whether that design is consistent with reasonably foreseeable automation and materials processing technologies. These technologies need not necessarily be well established or even state-of-the-art, but should at least be conceivable in the context of a dedicated R&D effort spanning the next two decades. It is interesting to note that computer programs capable of self-replication have been written in many different programming languages (Burger et al., 1980; Hay, 1980), and that simple physical machines able to replicate themselves in highly specialized environments have already been designed and constructed (Jacobson, 1958;Morowitz, 1959;Penrose,1959).
Another major requirement for concept credibility is a plausible system configuration. Proposed designs for selfreplicating systems (SRS) must be sufficiently detailed to permit the generation of work breakdown structures, subsystem operational flowcharts, mass and energy throughput calculations, and at least preliminary closure (see sec.5.3.6) analyses.
A related requirement is plausible mission scenarios. Research and development costs for the proposed design should be many orders of magnitude less than the Gross National Product. The mission must not require launch and support facilities which cannot or will not be available in the next two or three decades. The mission must entail reasonable flight times, system lifetimes, growth rates, production rates, and so forth. The problems of reliability and repair should be addressed.
The final requirement for concept credibility is positive societal impact. A given SRS design must be economically, politically, and socially feasible, or else it may never be translated into reality even if the technology to do so exists. A general discussion of the implications of replicating systems appears in section 5.5, but the team has arrived at no firm conclusions regarding concept feasibility in this area. More research is clearly required.
5.3.2 Concept Definition
In order to demonstrate SRS concept credibility, specific system designs and mission scenarios must be subjected to a detailed feasibility analysis. The first step in this process is to conceptualize the notion of replicating systems in as broad an engineering context as possible. Many kinds of replicating machine systems have been proposed and considered during the course of the study. Some of these place emphasis on different types of behavior than others.
Consider a "unit machine" which is the automata equivalent of the atom in chemistry or the cell in biology—the smallest working system able to execute a desired function and which cannot be further subdivided without causing loss of that function. The unit machine may be comprised of a number of subunits, say, A, B, C, and D. These subunits may be visualized in terms of structural descriptions (girders, gearboxes, generators), functional descriptions (materials processing, parts fabrication, mining, parts assembly), or any other complete subset-level descriptions of the entire system.
SRS may be capable of at least five broad classes of machine behavior:
Production—Generation of useful output from useful input. The unit machine remains unchanged in the process. This is a "primitive" behavior exhibited by all working machines including replicating systems.
Replication - Complete manufacture of a physical copy of the original unit machine, by the unit machine.
Growth—Increase in mass of the original unit machine by its own actions, retaining the physical integrity of the original design.
EVOLUTION REPLICATION PRODUCTION
Figure 5.5.- Five basic classes of SRS behavior.
Evolution—Increase in complexity of structure or function of the unit machine, by adding to, subtracting from, or changing the character of existing system subunits.
Repair—Any operation performed by a unit machine upon itself, which does not alter unit population, designed unit mass, or unit complexity. Includes reconstruction, reconfiguration, or replacement of existing subunits.
These five basic classes of SRS behavior are illustrated in figure 5.5.
Replicating systems, in principle, may be designed which can exhibit any or all of these machine behaviors. In actual practice, however, it is likely that a given SRS format will emphasize one or more kinds of behaviors even if capable of displaying all of them. The team has considered two specific replicating systems designs in some detail. The first (cf. von Tiesenhausen and Darbro, 1980), which may be characterized as a unit replication system, is described in section 5.3.3. The second (cf. Freitas, 1980a; Freitas and Zachary, 1981), which can be characterized as a unit growth system, is outlined in section 5.3.4. The team decided to concentrate on the possibility of fully autonomous or "unmanned" SRS, both because these are more challenging from a technical standpoint than either manned or teleoperated systems and also because the latter has already been lredted to some degree elsewhere in this report (see chap. 4).
R EPAIR RECONFIGURATION REPLACEMENT
5.3.3 Unit Replication: A Self-Replicating System Design
The SRS design for unit replication is intended to be a fully autonomous, general-purpose self-replicating factory to be deployed on the surface of planetary bodies or moons. The anatomy of an SRS is defined by two end conditions: (I) the type and quantity of products required within a certain time, and (2) the available material needed to manufacture these products as well as the SRS itself.
There are four major subsystems which comprise each SRS unit, as shown in figure 5.6. First, a materials processing subsystem acquires raw materials from the environment and prepares industrial feedstock from these substances. Second, a parts production subsystem uses the feedstock to make machines or other parts. At this point SRS output may take two forms. Parts may flow to the universal constructor subsystem, where they are used to construct a new SRS (replication). Or, parts may flow to a production facility subsystem to be made into commercially useful products. The SRS also has a number of other important but subsidiary subsystems, including a materials depot, parts depots, product depot, control and command, and an energy system.
The work breakdown structure given in figure 5.7 lists all SRS elements studied, and each is briefly described below.
REPLICATION MATERIAL FLOW REPLICATION
MP = MATERIALS PROCESSING
PP = PARTS PRODUCTION
PF = PRODUCTION FACILITY
UC = UNIVERSAL CONSTRUCTOR
Materials processing and feedstock production. In this system, raw materials are gathered by strip or deep milling. They are then analyzed, separated, and processed into industrial feedstock components such as sheets, bars, ingots, castings, and so forth, which are laid out and stored in the materials depot. The processing subsystem has a high degree of autonomy including self-maintenance and repair. It is linked to a central supervisory control system (see below).
The materials processing subsystem is shown schematically in figure 5.8.
Materials depot. The materials depot collects and deposits in proper storage locations the various feedstock categories according to a predetermined plan. This plan ensures that the subsequent fabrication of parts proceeds in the most efficient and expeditious manner possible. The depot also serves as a buffer during interruptions in normal operations caused by failures in either the materials processing subsystem (depot input) or in the parts production subsystem (at depot output).
Parts production plant. The parts production plant selects and transports industrial feedstock from the materials depot into the plaet, thee (6<i>~tAs %11 p1tt I~ qBilG } for SRS production or replication activities. Finished parts are stored in the production parts and the replication parts depots, respectively. The parts production plant is highly automated in materials transport and in distribution, production, control, and subassembly operations.
Figure 5. 6. - Functional schematic of unit replication SRS.
Figure 5. 7.- Work breakdown structure for SRS.
The parts production plant subsystem is shown schematically in figure 5.9.
Parts depots. There are two parts depots in the present design. These are called the production parts depot and the replication parts depot.
Parts are stored in the production parts depot exclusively for use in the manufacture of useful products in the production facility. If certain raw materials other than parts and subassemblies are required for production, these materials are simply passed from the materials depot through the parts production plant unchanged. The parts production depot also acts as a buffer during interruptions in normal operations caused by temporary failures in either the parts production plant or the production facility.
Parts and subassemblies are stored in the replication parts depot exclusively for use in the replication of complete SRS units. Storage is in lots earmarked for specific facility construction sites. The replication parts depot also serves as buffer during interruptions in parts production plant or universal constructor operations.
Figure 5. 7.- Concluded.
Production facility. The production facility manufactures the desired products. Parts and subassemblies are picked up at the production parts depot and are transported to the production facility to be assembled into specific useful products. Finished products are then stored in the products depot. Ultimately these are collected by the product retrieval system for outshipment.
Universal constructor. The universal constructor manufactures complete SRS units which are exact duplicates of the original system. Each replica can then, in turn, construct more replicas of itself, and so on. The universal constructor retains overall control and command responsibility for its own SRS as well as its replicas, until the control and command functions have also been replicated and transferred to the replicas. These functions can be overridden at any time by external means.
The universal constructor subsystem consists of two major, separate elements — the stationary universal constructor (fig. 5.10) and the mobile universal constructors (fig. 5.11). This composite subsystem must successfully perform a number of fundamental tasks, including receiving, sorting, loading, and transporting parts and subassemblies; assembling, constructing, installing, integrating, and testing SRS systems; starting and controlling SRS operations; and copying and transferring instructions between system components.
Figure 5.8. - SRS materials processing subsystem.
Products depot. The outputs of the production facility are stored in the products depot, ready for retrieval. Major hardware components are neatly stacked for ready access by the product retrieval system. Consumables such as elemental oxygen are stored in reusable containers that are returned empty to the production facility. The products depot also serves as a buffer against variable output and retrieval rates.
Product retrieval system. The product retrieval system collects the outputs of all SRS units in an "SRS field" and carries them to an outside distribution point for immediate use or for subsequent outshipment. The dashed lines in figure 5.11 indicate one possible solution to this problem in a typical SRS field. Other solutions are possible—careful consideration must be given to SRS field configuration to arrive at an optimum product retrieval system design.
Command and control systems. The master control and command system, located within the stationary universal constructor, is programmed to supervise the total SRS operation and to communicate both with the peripheral controls of the mobile universal constructors during the selfreplication phase and with the replicated stationary universal constructor during the transfer of command and control for the operation of the new SRS unit.
(fig. 5.11). This composite subsystem must successfully perform a number of fundamental tasks, including receiving, sorting, loading, and transporting parts and subassemblies; assembling, constructing, installing, integrating, and testing SRS systems; starting and controlling SRS operations; and copying and transferring instructions between system components.
REFERENCE: "LUNAR STRIP MINING SYSTEM" W. DAVID CARRIER , JUNE 1978
Figure 5.8. - SRS materials processing subsystem.
Products depot. The outputs of the production facility are stored in the products depot, ready for retrieval. Major hardware components are neatly stacked for ready access by the product retrieval system. Consumables such as elemental oxygen are stored in reusable containers that are returned empty to the production facility. The products depot also serves as a buffer against variable output and retrieval rates.
Product retrieval system. The product retrieval system collects the outputs of all SRS units in an "SRS field" and carries them to an outside distribution point for immediate use or for subsequent outshipment. The dashed lines in figure 5.11 indicate one possible solution to this problem in a typical SRS field. Other solutions are possible - careful consideration must be given to SRS field configuration to arrive at an optimum product retrieval system design.
Command and control systems. The master control and command system, located within the stationary universal constructor, is programmed to supervise the total SRS operation and to communicate both with the peripheral controls of the mobile universal constructors during the selfreplication phase and with the replicated stationary universal constructor during the transfer of command and control for the operation of the new SRS unit.
Figure 5.9.- SRS parts production plant subsystem.
STORAGE PRODUCTS COMMAND RECEIVER
Figure 5.10. - SRS stationary universal constructor.
The master control and command system operates its own SRS unit through individual communication links which address the local control and command systems of individual SRS elements. In this way the master control and command system supervises the condition and operations of its own system elements, from materials acquisition through end product retrieval.
Energy system. The power requirements for the present design may be in gigawatt range. Hence, a single energy source (such as a nuclear power plant) would be excessively massive, and would be difficult to replicate in any case. This leaves solar energy as the lone viable alternative. Daylight options include: (1) central photovoltaic with a ground cable network, (2) distributed photovoltaic with local distribution system, (3) individual photovoltaic, and (4) satellite power system, with microwave or laser power transmission to central, local, or individual receivers. Nighttime power options include MHD, thermionics, or turbogenerators using fuel generated with excess capacity during daytime. Oxygen plus aluminum, magnesium, or calcium could be used for fuel. A 155to efficient central silicon photovoltaic power station has been assumed in the reference design, with an output of tens of gigawatts and a size on the order of tens of square kilometers.
Figure 5.11.- SRS mobile universal constructors.
Each SRS produces, in addition to its scheduled line of regular products, a part of the photovoltaic energy system equal to the energy needs of its replicas. These are retrieved along with the regular products by the product retrieval system and are assembled on-site to increase energy system capacity according to demand during the self-replication phase.
SRS deployment and expansion. A complete SRS factory unit, erected on the surface of the Moon, might appear as illustrated in figure 5.12.
As a unit replication scheme, the multiplication of SRS units proceeds from a single primary system to many hundreds of replica systems. This expansion must be carefully planned to reach the desired factory output capacity without running out of space and materials. Figure 5.13 shows one possible detailed growth plan for the geometry of an SRS field. In this plan, each SRS constructs just three replicas, simultaneously, then abandons replication and goes into full production of useful output. After the three generations depicted, an SRS field factory network 40 units strong is busy manufacturing products for outshipment.
The routes taken by mobile universal constructors are shown as solid lines, the product retrieval routes as dashed lines.
Figure 5.14 shows another possible expansion geometry. Again, each SRS constructs just three replicas, but sequentially rather than simultaneously. The end result is a field of 326 individual units after nine cycles of replication. Output is collected by the product retrieval system and taken to an end product assembly/collection system where end products undergo final assembly and other operations preparatory to outshipment. A more detailed discussion of expansion scenarios for SRS fields may be found in von Tiesenhausen and Darbro (1980).
Proposed development and demonstration scenario. It is proposed that the practical difficulties of machine replication should be confronted directly and promptly by a dedicated development and demonstration program having four distinct phases.
In Phase A of the development scenario, a robot manipulator will be programmed to construct a duplicate of itself from supplied parts and subassemblies. The original robot then makes a copy of its own operating program and inserts this into the replica, then turns it on, thus completing the duplication process (see appendix 5J). To complete Phase A; the replica must construct a replica of itself, repeating in every way the actions of the original robot. The rationale for the second construction, called the Fertility Test, is to demonstrate that the capacity for self-replication has in fact been transmitted from parent machine to offspring.
Figure 5.12.- Self-replicating lunar factory.
In Phase B of the development and demonstration scenario, the robot manipulator will be supplied with numerous additional parts so it can assemble objects of interest other than replicas of itself. This is intended to show that the system is able to construct useful products in addition to the line of robot duplicates.
In Phase C the manipulator system is still required to construct replicas and useful products. However, the robot now will be supplied only with industrial feedstock such as metal ingots, bars, and sheets, and must fabricate all necessary parts and subassemblies on its own. Successful completion of Phase C is expected to be much more difficult than the two earlier phases. The reason is that the parts fabrication machines must themselves be constructed by the robot manipulator and, in addition, all parts and subassemblies comprising the newly introduced fabrication machines must also be made available to the manipulator. Fabricator machines thus must be programmed to make not only the parts required for robot manipulators and useful products, but also their own parts and subassemblies as well. This raises the issue of parts closure, a matter which is discussed in section 5.3.6.
In Phase D, the system developed in the previous phase is retained with the exception that only minerals, ores, and soils of the kind naturally occurring on terrestrial or lunar surfaces are provided. In addition to all Phase C capabilities, the Phase D system must be able to prepare industrial feedstock for input to the fabrication machines. Successful completion of Phase D is expected to be the most difficult of all because, in addition to the parts closure problem represented by the addition of materials processing machines, all chemical elements, process chemicals, and alloys necessary for system construction and operation must be extracted and prepared by the materials processing machines. This raises the issue of materials closure (see also sec. 5.3.6). The completion of Phase D will yield an automatic manufacturing facility which, beginning with "natural" substrate, can replicate itself.
Figure 5.13. - Possible growth plan with simultaneous replica construction, suitable for geometry of an SRS field.
Figure 5.14.- SRS growth plan with sequential replication.
This progressive development of a replicating factory will serve to verify concept feasibility, clarify the functional requirements of such a system, and identify specific technological problem areas where additional research in automation and robotics is needed. A minimum demonstration program should be designed to gain engineering under standing, confidence, and hands-on experience in the design A and operation of replicating systems. (See sec. 5.6.) The question of when the results of an Earth-based development and demonstration project should be translated to lunar requirements, designs, and construction remains open. On the one hand, it may be deemed most practical to complete Phase D before attempting a translation to a design better suited to a lunar or orbital environment. On the other hand, major system components for a lunar facility undoubtedly could be undertaken profitably earlier in concert with Phase C and D development. The proposed development and demonstration scenario is described in greater detail in von Tiesenhausen and Darbro (1980).
5.3.4 Unit Growth: A Growing Lunar Manufacturing Facility
The Lunar Manufacturing Facility (LMF) demonstrating SRS unit growth is intended as a fully automatic general purpose factory which expands to some predetermined adult size starting from a relatively tiny "seed" initially deposited on the lunar surface. This seed, once deployed on the Moon, is circular in shape, thus providing the smallest possible perimeter/surface area ratio and minimizing interior transport distances. Expansion is radially outward with an accelerating radius during the growth phase. Original seed mass is 100 tons.
The replicating LMF design encompasses eight fundamental subsystems. Three subsystems are external to the main factory (transponder network, paving, and mining robots). The LMF platform is divided into two identical halves, each comprised of three major production subsystems: (I) the chemical processing sector accepts raw lunar materials, extracts needed elements, and prepares process chemicals and refractories for factory use; (2) the fabrication sector converts these substances into manufactured parts, tools, and electronics components; and (3) the assembly sector, which assembles fabricated parts into complex working machines or useful products of any conceivable design. (Each sector must grow at the same relative rate for uniform and efficient perimeter expansion.) Computer facilities and the energy plant are the two remaining major subsystems. (See fig. 5.15 .)
Transponder network. A transponder network operating in the gigahertz range assists mobile LMF robots in accurately fixing their position relative to the main factory complex while they are away from it. The network, described briefly in appendix SB, is comprised of a number of navigation and communication relay stations set up in a well defined regular grid pattern around the initial seed and the growing LMF complex.
Figure 5.15.- Functional schematic of unit growth SRS.
Paving robots. In order to secure a firm foundation upon which to erect seed (and later LMF) machinery, a platform of adjoining flat cast basalt slabs is required in the baseline design. A team of five paving robots lays down this foundation in a regular checkerboard pattern, using focused solar energy to melt pregraded lunar soil in situ. (See app. 5C.)
Mining robots. As described in appendix SD, LMF mining robots perform six distinct functions in normal operation: (1) strip mining, (2) hauling, (3) landfilling, (4) grading, (5) cellar-digging, and (6) towing. Lunar soil is stripmined in a circular pit surrounding the growing LMF. This material is hauled back to the factory for processing, after which the unused slag is returned to the inside edge of the annular pit and used for landfill which may later be paved over to permit additional LMF radial expansion. Paving operations require a well graded surface, and cellar digging is necessary so that the LMF computer may be partially buried a short distance beneath the surface to afford better protection from potentially disabling radiation and particle impacts. Towing is needed for general surface transport and rescue operations to be performed by the mining robots. The robot design selected is a modified front loader with combination roll-back bucket/dozer blade and a capacity for aft attachments including a grading blade, towing platform, and a tow bar.
Chemical processing sectors. Mining robots deliver raw lunar soil strip-mined at the pit into large input hoppers arranged along the edge of entry corridors leading into the chemical processing sectors in either half of the LMF. This material is electrophoretically separated (Dunning and Snyder, 1981; see sec. 4.2.2) into pure minerals or workable mixtures of minerals, then processed using the HF acid-leach method (Arnold et al., 1981; Waldron et al., 1979) and other specialized techniques to recover volatiles, refractories, metals, and nonmetallic elements. Useless residue and wastes are collected in large output hoppers for landfill. Buffer storage of materials output is on site. Chemical processing operations are shown schematically in figure 5.16, and are detailed in appendix SE.
Fabrication sectors. The LMF fabrication sector outlined in appendix SF is an integrated system for the production of finished aluminum or magnesium parts, wire stock, cast basalt parts, iron or steel parts, refractories, and electronics parts. Excepting electronics (Zachary, 1981) there are two major subsystems: (1) the casting subsystem, consisting of a casting robot to make molds, mixing and alloying furnaces for basalt and metals, and automatic molding machines to manufacture parts to low tolerance using the molds and alloys prepared; and (2) the laser machining and finishing subsystem, which performs final cutting and machining of various complex or very-close-tolerance parts. The basic operational flowohart for parts fabrication is shown in figure 5.17.
Assembly sectors. Finished parts flow into the automated assembly system warehouse, where they are stored and retrieved by warehouse robots as required. This subsystem provides a buffer against system slowdowns or temporary interruptions in service during unforeseen circumstances. The automated assembly subsystem requisitions necessary parts from the warehouse and fits them together to make subassemblies which are inspected for structural and functional integrity. Subassemblies may be returned to the warehouse for storage, or passed to the mobile assembly and repair robots for transport to the LMF perimeter, either for internal repairs or to be incorporated into working machines and automated subsystems which themselves may contribute to further growth. The basic operational flowchart for SRS parts assembly is shown in figure 5.18, and a more detailed presentation may be found in appendix SG.
Computer control and communications. The seed computers must be capable of deploying and operating a highly complex, completely autonomous factory system. The original computer must erect an automated production facility, and must be expandable in- order to retain control as the LMF grows to its full "adult" size. The computer control subsystem coordinates all aspects of production, scheduling, operations, repairs, inspections, maintenance, and reporting, and must stand ready to respond instantly to emergencies and other unexpected events. Computer control is nominally located at the hub of the expanding LMF disk, and commands in hierarchical fashion a distributed information processing system with sector computers at each node and sector subsystems at the next hierarchical level of control. Communications channels include the transponder network, direct data bus links, and E2ROM messenger chips (firmware) for large data block transfers.
Using ideas borrowed from current industrial practice, top-down structured programming, and biology, Cliff (1981) has devised a system architecture which could perform automated design, fabrication, and repair of complex systems. This architecture, presented in appendix SH, is amenable to straightforward mathematical analysis and should be a highly useful component of the proposed lunar SRS. Further work in this area should probably include a survey of industrial systems management techniques (Carson, 1959) and the theory of control and analysis of large-scale systems (Sandell et al., 1978).
Figure 5.16.- LMF chemical processing sector. Operations.
In a practical sense, it is quite possible to imagine the lunar SRS operating nonautonomously (Johnsen, 1972). For instance, the in situ computer could be used simply as a teleoperation-management system for operations controlled directly by Earth-based workers. Material factory replication would proceed, but information necessary to accomplish this would be supplied from outside. An intermediate alternative would permit the on-site computer to handle mundane tasks and normal functions with humans retaining a higher-level supervisory role. Yet another possibility is that people might actually inhabit the machine factory and help it reproduce—manned machine economies can also self-replicate.
Solar canopy. The solar canopy is a "roof" of photovoltaic solar cells, suspended on a relatively flimsy support web of wires, crossbeams and columns perhaps 3-4 m above ground level. The canopy covers the entire LMF platform area and expands outward as the rest of the facility grows. The solar canopy and power grid provide all electrical power for LMF systems. Canopy components may be stationary or may track solar motions using heliostats if greater eff1ciency is required. A further discussion of canopy design and rationale may be found in appendix 5I.
Mass, power, and information requirements. Seed subsystem masses and power requirements scale according to the total system mass assumed. SRS can be reduced indefinitely in size until its components begin to scale nonlinearly. Once this physical or technological limit is reached for any subsystem component, comprehensive redesign of the entire factory may become necessary.
Figure 5.18.- LMF assembly sector: Operations.
A seed mass of 100 tons was selected in the present study for a number of reasons. First, 100 tons is a credible system mass in terms of foreseeable NASA launch capabilities to the lunar surface, representing very roughly the lunar payload capacity of four Apollo missions to the Moon. Second, after performing the exercise of specifying seed components in some detail it is found that many subsystems are already approaching a nonlinear scaling regime for a 100-ton LMF. For instance, according to Criswell (1980, private communication) the minimum feasible size for a linear-scaling benchtop HF acid-leach plant for materials processing is about 1000 kg; in the present design, two such plants are required with a mass of 1250 kg each. Third, the results of a previous study (Freitas, 1980a) which argued the feasibility of 433-ton seed in the context of an interstellar mission (inherently far more challenging than a lunar factory mission) were compared with preliminary estimates of 15-107 tons for partially self-replicating lunar factories of several different types (O'Neill et al., 1980), and an intermediate trial value of 100 tons selected. The 100-ton figure has appeared in numerous public statements by former NASA Administrator Dr. Robert A. Frosch (lecture delivered at Commonwealth Club, San Francisco, Calif., 1979, and personal communication,1980) and by others in prior studies (Bekey and Naugle, 1980; Giacconi et al., working paper of the Telefactors Working Group, Woods Hole New Directions Workshop, 1979). Finally, it was decided to use a specific system mass rather than unscaled relative component mass fractions to help develop intuitive understanding of a novel concept which has not been extensively studied before.
For reasons similar to the above, an SRS strawman replication time of 1 year was taken as appropriate. The ranges given in table 5.1, drawn from the analysis presented in appendixes 5B-51, are estimates of the mass and power requirements of an initial seed system able to manufacture 100 tons of all of its own components per working year, hence, to self-replicate. These figures are consistent with the original estimate of a 100 ton circular LMF seed with an initial deployed diameter of 120 m, so feasibility has been at least tentatively demonstrated. However, it must be emphasized that the LMF seed design outlined above is intended primarily as a proof of principle. Numerical values for system components are only crude estimates of what ultimately must become a very complex and exacting design.
Information processing and storage requirements also have been collected and summarized in table 5.1, and lie within the state-of-the-art or foreseeable computer technologies. These calculations, though only rough approximations, quite likely overestimate real needs significantly because of the conservative nature of the assumptions employed. (See also sec.5.2.3.)
SRS mission overview. In the most general case of fully autonomous operation, a typical LMF deployment scenario might involve the following initial sequence:
(1) The predetermined lunar landing site is mapped from orbit to l-m resolution across the entire target ellipse.
(2) Seed lands on the Moon, as close to dead center of the mapped target area as possible navigationally.
(3) Mobile assembly and repair robots, assisted by mining robots, emerge from the landing pod and erect a small provisional solar array to provide interim power until the solar canopy is completed.
(4) LMF robots, with the computer, select the precise site where erection of the original seed will commence. This decision will already largely have been made based on orbital mapping data, but ground truth will help refine the estimate of the situation and adjust for unexpected variations.
(5) Mobile robots emplace the first three stations of the transponder network (the minimum necessary for triangulation), calibrate them carefully, and verify that the system is in good working order.
(6) Mining robots equipped with grading tools proceed to the construction site and level the local surface.
(7) Five paving robots disembark and begin laying down the seed platform in square grids. This requires one working year for completion.
(8) When a sufficiently large platform section has been completed, seed mobile robots transfer the main computer to a place prepared for it at the center of the expanding platforrn disk.
(9) Erection of the solar canopy begins, followed by each of the seed sectors in turn, starting with the chemical processing. Total time to unpack the landing pod after moonfall is one working year, conducted in parallel with paving and other activities. The completed seed factory unit, unfurled to a 120 m diam on the surface of the Moon 1 vear after landing, might appear as shown in figure 5.19.
TABLE 5.1.—SEED MASS AND POWER REQUIREMENTS ESTIMATES
Seed subsystem
Transponder network
Paving robots
Mining robots
Chemical processing sector (S)
Fabrication sector (S) Electronics Floor map Totals
Assembly sector (S) Assembly robots Warehouse subsystem Floor map
Automated transport vehicles
Mobile assembly and repair robots
Computer central orbital site map
Solar canopy
Totals
Estimated mass of
100 ton/yr seed,
1,000
12,000
4,400
15,300-76,400
3,000)
Estimated power of 100 ton/yr seed, W
Up to 104
Up to 104
380,000-11,000,000
137-20,400 270-345,000
83-1,150 83-19,600
1,000 10,000
1,000
4,000 2,200
22,000
63,100-145,600
Computer Computer processor, memory, bits to operate hitc tr, {1>coriho
105 ?
l-lOX 106
4-7X 108
9.4X 107
lolo
109
07
6,000 107
40,000 4X 109
37.000 (1.6X 101 °)
Nomirial annual seed output 100,000 1.7 MW
The LMF has two primary operational phases—growth and production. The optimal program would probably be to "bootstrap" (grow) up to a production capacity matching current demand, then reconfigure for production until demand increases, thus necessitating yet further growth (O'Neill et al., 1980). Growth and production of useful output may proceed sequentially, cyclically, or simultaneously, though the former is preferred if large subsystems of the lunar factory must be reconfigured to accommodate the change.
The LMF also may exhibit replicative behavior if and when necessary. Replicas of the original seed could be constructed much like regular products and dispatched to remote areas, either to increase the total area easily subject to utilization or to avoid mortality due to depletion of local resources or physical catastrophes. The scheduling of factory operational phases is very flexible,-as shown schematically in figure 5.20, and should be optimized for each mission and each intended use.
5.3.5 Lunar SRS Growth and Productivity
As the study progressed, the team noted a developing convergence between the two designs for SRS described in sections 5.3.3 and 5.3.4. Both require three major subsystems—materials processing, fabrication, and assembly plus a variety of support systems, and each is capable of replication and useful production. Both display exponential expansion patterns.
Figure 5.19.- Self-growing lunar factory.
Figure 5.20. - Flexible scheduling of LMF operational phases.
Of course, in a finite environment exponential growth cannot continue indefinitely. Geometrical arguments by - Taneja and Walsh (1980, Summer Study document) suggest that planar packing of triangular, cubic, or hexagonal units can expand exponentially only for as many generations as each unit has sides, assuming that once all sides are used up no further doubling can occur by the enclosed unit.
Growth is quadratic from that time on. However, in real physical systems such as the developing LMF, enclosure need not preclude material communication with exterior units. Selected ramification of communication, control, and materials transportation channels or internal component rearrangement, reconfiguration, or specialization can prevent "starvation" in the inner regions of the expanding system. Hence, SRS exponential growth may continue until limited either by purposeful design or by the - specific configuration of the external environment. Assuming that a lO0-ton seed produces 100 tons/year of the same materials of which it is composed, then if T is elapsed time and N is number of seed units or seed mass-equivalents generated during this time, T = I + log2 N for simple exponential "doubling" growth. (There is no replication in the first year, the time required for initial setup.) If P is productivity in tons/year, then P = 100 log2 N.; However, the above is valid only if each unit works only on its own replica. If two or more units cooperate in the construction of a single replica, still more rapid "fast exponential" growth is possible. This is because new complete replicas or LMF subsystems are brought on line sooner, and thence may begin contributing to the exponentiation earlier than before. Using the above notation, the "fast exponential" growth rate is given by T= 1 + 1/2 + + 1/N in the optimum case where all available machines contribute directly to the production of the next unit.
Growth rates and productivities are tabulated for exponential and "fast-exponential" expansion in table 5.2. Note that in just 10 years the output of such a facility could grow to approximately one million tons per year. If allowed to expand for 18 years without diversion to production, the factory output could exponentiate to more than 4 X 109 tons per year, roughly the entire annual industrial output of all human civilization.
Useful SRS products may include lunar soil thrown into orbit by mass drivers for orbital processing, construction projects, reaction mass for deep space missions, or as radiation shielding; processed chemicals and elements, such as oxygen to be used in space habitats, as fuel for interorbital vehicles, and as reaction mass for ion thrusters and mass drivers; metals and other feedstock ready-made for space construction or large orbital facilities for human occupation (scientific, commercial, recreational, and medical); components for large deep-space research vessels, radio telescopes, and large high-power satellites; complex devices such as machine shop equipment, integrated circuits, sophisticated electronics gear, or even autonomous robots, teleoperators, or any of their subassemblies; and solar cells, rocket fuels, solar sails, and mass driver subassemblies. Also, a 100-ton seed which has undergone thousand-fold growth or replication represents a 2 GW power generating capacity, plus a computer facility with a 16,000 Gbit processing capability and a total memory capacity of 272,000 Gbits. These should have many useful applications in both terrestrial and space industry.
5.3.6 Closure in Self-ReplicatingSystems
Fundamental to the problem of designing self-replicating systems is the issue of closure.
TABLE 5.2.—GROWTH RATES AND PRODUCTIVITY FOR EXPONENTIAL SRS EXPANSION
"Fast-exponential" growth, T = 1 yr
(About 3 billion seed units would completely cover the entire lunar surface)
In its broadest sense, this issue reduces to the following question: Does system function (e.g., factory output) equal or exceed system structure (e.g., factory components or input needs)? If the answer is negative, the system cannot independently fully replicate itself; if positive, such replication may be possible.
Consider, for example, the problem of parts closure. Imagine that the entire factory and all of its machines are broken down into their component parts. If the original factory cannot fabricate every one of these items, then parts closure does not exist and the system is not fully self-replicating .
In an arbitrary system there are three basic requirements to achieve closure:
(1) Matter closure - can the system manipulate matter in all ways necessary for complete self-construction?
(2) Energy closure—can the system generate sufficient energy and in the proper format to power the processes of self-construction?
(3) Information closure can the system successfully command and control all processes required for complete self-construction?
System productivity, tons/yr
100 400 1 ,100 3,100 8,300
22,700 61,600 167,400 455,000 1,236,700 3,361,700 9,138,000 24,839,800 67,521,500
183,542,600 498,920,500 1,356,206,600 3,686,551,700 10,021,086,500 27,240,137,200
(~2 km-wide asteroid/yr)
Partial closure results in a system which is only partially self-replicating. Some vital matter, energy, or information must be provided from the outside or the machine system will fail to reproduce. For instance, various preliminary studies of the matter closure problem in connection with the possibility of "bootstrapping" in space manufacturing have concluded that 90-96% closure is attainable in specific nonreplicating production applications (Bock, 1979; Miller and Smith, 1979; O'Neill et al., 1980). The 4-10% that still must be supplied sometimes are called "vitamin parts." These might include hard-to-manufacture but lightweight items such as microelectronics components, ball bearings, precision instruments and others which may not be costeffective to produce via automation off-Earth except in the longer term. To take another example, partial information closure would imply that factory-directive control or supervision is provided from the outside, perhaps (in the case of a lunar facility) from Earth-based computers programmed with human-supervised expert systems or from manned remote teleoperation control stations on Earth or in low Earth orbit.
The fraction of total necessary resources that must be supplied by some external agency has been dubbed the "Tukey Ratio" (Heer, 1980). Originally intended simply as an informal measure of basic materials closure, the most logical form of the Tukey Ratio is computed by dividing the mass of the external supplies per unit time interval by the total mass of all inputs necessary to achieve self-replication. (This is actually the inverse of the original version of the ratio.) In a fully self-replicating system with no external inputs, the Tukey Ratio thus would be zero (0%).
It has been pointed out that if a system is "truly isolated in the thermodynamic sense and also perhaps in a more absolute sense (no exchange of information with the environment) then it cannot be self-replicating without violating the laws of thermodynamics" (Heer,1980). While this is true, it should be noted that a system which achieves complete "closure" is not "closed" or "isolated" in the classical sense. Materials, energy, and information still flow into the system which is thermodynamically "open"; these flows are of indigenous origin and may be managed autonomously by the SRS itself without need for direct human intervention.
Closure theory. For replicating machine systems, complete closure is theoretically quite plausible; no fundamental or logical impossibilities have yet been identified. Indeed, in many areas automata theory already provides relatively unambiguous conclusions. For example, the theoretical capability of machines to perform "universal computation" and "universal construction" can be demonstrated with mathematical rigor (Turing, 1936; von Neumann, 1966; see also sec. 5.2), so parts assembly closure is certainly theoretically possible.
An approach to the problem of closure in real engineering-systems is to begin with the issue of parts closure by asking the question: can a set of machines produce all of its elements? If the manufacture of each part requires, on average, the addition of >1 new parts to product it, then an infinite number of parts are required in the initial system and complete closure cannot be achieved. On the other hand, if the mean number of new parts per original part is <1, then the design sequence converges to some finite ensemble of elements and bounded replication becomes possible.
The central theoretical issue is: can a real machine system itself produce and assemble all the kinds of parts of which it is comprised? In our generalized terrestrial indus-trial economy manned by humans the answer clearly is yes, since "the set of machines which make all other machines is a subset of the set of all machines" (Freitas et al.,1981). In space a few percent of total system mass could feasibly be supplied from Earth-based manufacturers as "vitamin parts." Alternatively, the system could be designed with components of very limited complexity (Heer, 1980). The minimum size of a self-sufficient "machine economy" remains unknown.
Von Tiesenhausen and Darbro (1980) similarly argue that a finite set of machines can produce any machine element . Their reasoning, outlined in figure 5.21, is as follows:
(1) If all existing machines were disassembled into their individual parts there would obviously be a finite number of parts, many of them identical, and a large number would be of common categories like shafts, motors, wiring, etc. The only differences between the machines would be a different selection, different arrangement, and different dimensions of this finite number of parts.
(2) A finite number of parts involves a finite number of machine operations, this number being less than the number of parts because some machines can make more than one kind of parts.
(3) Therefore, the number of machines is finite and less than the number of operations.
This reasoning can then be generalized to say: "Every existing machine can be reduced to a finite set of machine elements, and there exists a finite set of machine operations." (Still, of course, a limited number of standard elements should be developed and machine operations limited as much as practical by substitution, in order to minimize the number of parts and machine operations.)
Figure 5.21.- Closure of SRS parts production.
Similar arguments may be applied to materials processing and feedstock production. There exists a finite number of different materials anywhere. There is a finite number of materials processes which is less than the number of materials because single processes result in various materials (e.g., silicon and oxygen). Hence, there is a finite number of materials processing robot systems needed for an SRS. Also, there is a flnite and rather limited number of feedstock requirements such as bars, rods, ingots, plates, etc. The number of materials is much less than the number of parts; therefore, a finite number of parts fabrication robots is required for an SRS.
Closure engineering In actual practice, the achievement of full closure will be a highly complicated, iterative engineering design process. Every factory system, subsystem, component structure, and input requirement (Miller and Smith, 1979) must be carefully matched against known factory output capabilities. Any gaps in the manufacturing flow must be filled by the introduction of additional machines, whose own construction and operation may create new gaps requiring the introduction of still more machines.
The team developed a simple iterative procedure for generating designs for engineering systems which display complete closure. The procedure must be cumulatively iterated, first to achieve closure starting from some initial design, then again to eliminate overclosure to obtain an optimally efficient design. Each cycle is broken down into a succession of subiterations which ensure three additional dimensions of closure:
(1) Qualitative closure—can, say, all parts be made?
(2) Quantitative closure - can, say, enough parts be made?
(3) Throughput closure — can parts be made fast enough?
In addition, each subiteration sequence is further decomposed into design cycles for each factory subsystem or component, as shown in figure 5.22.
The procedure as outlined, though workable in theory, appears cumbersome. Further work should be done in an attempt to devise a more streamlined, elegant approach.
Quantitative materials closure - numerical results. In the context of materials processing, "closure" is a relationship between a given machine design and a given particular substrate from which the machine's elemental chemical constituents are to be drawn. Hence the numerical demonstration of closure requires a knowledge of the precise composition both of the intended base substrate to be utilized and of the products which the SRS must manufacture from that substrate. Following a method suggested by the work of Freitas (1980a), a modified "extraction ratio" Ry is defined as the mass of raw substrate material which must be processed (input stream) to obtain a unit mass of useful system output having the desired mass fraction of element n (output stream).
Consider the significance of the extraction ratio to the problem of materials closure. Assume that the final product is to be composed of elements x, y, and z. An Rx = 1 means that 1 kg of lunar soil contains exactly the mass of element x needed in the manufacture of 1 kg of the desired output product. On the other hand, Ry = 10 means that 10 kg of lunar regolith must be processed to extract all of element y required in 1 kg of final product. The difference between Rx and Ry may signify that y is more rare in lunar soil than x, or that the two elements are equally abundant but ten times more y than x is required (by weight) in the final product. When the output stream is identical to the machine processing system itself, then the system is manufacturing more of itself -- self-replicating—and the extraction ratio becomes an index of system materials closure on an element-by-element basis.
The total net extraction ratio R is some function of the individual extraction ratios Rns and depends on the methods of materials processing employed. At worst, if only one element is recovered from a given mass of input stream ("parallel processing"), then R is the sum of all Rn. At best, if the input stream is processed sequentially to extract all desired elements in the necessary amounts ("serial processing"), then R is driven solely by the Rn of the element most difficult to extract, say, element z. That is, R = (Rn)maX = Rz, which is always equal to or smaller than the sum of all Rn. As serial processing should dominate in the lunar factory the latter formula is assumed for purposes of the present calculations. Note that Rn can be less than 1 for individual elements, but for an entire machine systemR must always be greater than or equal tol .
As a general rule, a low value for R implies that the system is designed for low mass throughput rates and is built from relatively few different chemical elements. A high value of R implies that many more elements are necessary and that a higher mass throughput rate will be accommodated to obtain them.
The "closure" of a given output stream (product) relative to a specified input stream (substrate) is computed by treating R as an independent variable. If In is the concentration of element n in mineral form in the input stream of lunar soil (kg/kg), En is the efficiency of chemical extraction of pure element n from its mineral form which is present in lunar soil (kg/kg), and °n is the concentration of element n in the desired factory output stream (kg/kg), then Rn = On/EnIn. Closure Cn for each element is defined as the mass of pure element n available in a system with a total net extraction ratio R per unit mass of output stream. For any given element, if R > Rn then all pure element n needed is already available within the system. In this case, Cn = °n. On the other hand, if R < Rn then the choice of R is too low; all the pure element n needed cannot be recovered, and more lunar soil must be processed to make up the difference if 100% closure is to be achieved. In this case, Cn = On(R/Rn), since the closure deficit is measured by the ratio of the chosen R to the actual Rn of the given element (i.e., how much the factory has, divided by how much the factory actually needs). Total net system closure C is simply the sum of all Cn for all elements n required in the output stream of the SRS factory (Freitas and Zachary, 1981)
EXPLANATION OF SYMBOLS
PROCEDURAL QUERIES = Q
Q1a —CAN SYSTEM ASSEMBLE ALL MACHINES OF WHICH IT IS COMPRISED7
Q2a —CAN SYSTEM ASSEMBLE ALL SUBASSEMBLIES OF WHICH ITS MACHINES ARE COMPRISED?
Q3a —CAN SYSTEM MACHINE ALL PARTS OF WHICH ITS MACHINES ARE COMPRISED?
04a —CAN SYSTEM EXTRACT ALL MATERIALS OF WHICH ITS PARTS ARE COMPRISED?
Q5a —CAN SYSTEM MINE ALL MINERALS OF WHICH ITS MATERIALS ARE COMPRISED?
Q6a —CAN SYSTEM TRANSPORT ALL (MACHINES/SUBASSEMBLIES/PARTS/MATERIALS/ MINERALS} OF WHICH IT IS COMPRISED?
Q7a —CAN SYSTEM VERI FY ALL {MACHINES/SUBASSEMBLIES/PARTS/MATERIALS/ MINERALS} OF WHICH IT IS COMPRISED?
Q8a —CAN SYSTEM WAREHOUSE ALL (MACHINES/SUBASSEMBLIES/PARTS/MATERIALS/ MINERALS} OF WHICH IT IS COMPRISED?
Q9a —CAN SYSTEM REPAIR ALL SUBSYSTEMS OF WHICH IT IS COMPRISED?
Q10e —CAN SYSTEM COMPUTERS CONTROL ALL SUBSYSTEMS OF WHICH IT IS COMPRISED?
Q1 1a —CAN SYSTEM ENERGY PLANT ENERGIZE ALL SUBSYSTEMS OF WHICH IT {THE SYSTEM} IS COMPRISED?
Qlb —CAN SYSTEM ASSEMBLE ENOUGH MACHINES TO REPLICATE THE ORIGINAL SYSTEM?
Q2b —CAN SYSTEM ASSEMBLE ENOUGH SUBASSEMBLIES TO REPLICATE THE ORIGINAL SYSTEM?
OUTPUT
SYSTEM COMPONENTS LISTS = C
Cla,b,c C2a,b,c C3a,b,c C4a,b,c C5a,b,c C6a,b,c C7a,b,c C&, b, c C9a,b,c ClOa,b,c C 11s, b, c
—MACHINES
—SUBASSEMBLIES
—PARTS
—MATERIALS
—MINERALS (SUBSTRATE}
—TRANSPORTATION FACILITIES
—VERIFICATION FACILITIES
—STORAGE FACILITIES
—REPAIR FACILITIES
—COMPUTER FACILITIES
—ENERGY FACILITIES
— CAN SYSTEM ENERGY PLANT PRODUCE ENOUGH ENERGY TO PERMIT REPLICATION OF THE ORIGINAL SYSTEM?
— CAN SYSTEM ASSEMBLE MACHINES FAST ENOUGH TO REPLICATE THE ORIGINAL SYSTEM WITHIN ESTABLISHED TIME CONSTRAINTS?
Figure 5.22.- Generalized closure engineering design cycles.
SYSTEM COMPONENTS ACTION LIST - A
Ala, b, c —NEW MACHINES
A2a, b, c —NEW SUBASSEMBLY-MAKING MACHINES
Alla, b, c —NEW ENERGY FACILITY-MAKING MACHINES
A#a —Add nrnv machines to make
A#b —Add nevs machines to make
A#c —increase replication time available or change machine design
To estimate the quantitative materials closure for the lunar SRS baseline designs proposed in sections 5.3.3 and 5.3.4, three different approaches were taken in an attempt to converge on a useful estimate of the composition of the output stream necessary for LMF selfreplication. First, the "seed" element distribution given by Freitas (1980a) in the context of a self-reproducing exploratory spaceprobe was adopted. These figures are derived from published data on the material consumption of the United States (the world's largest factory) during the years 1972-1976 (U.S. Bureau of Mines, 1978; U.S. Bureau of the Census, 1977, 1978). A second but less comprehensive measure called "demandite" is based on 1968 U.S. consumption data (Goeller and Weinberg, 1976). A molecule of "nonfuel demandite" is the average nonrenewable resource used by humans, less fuel resources (Waldron et al., 1979). Third, the direct estimate of LMF elemental composition presented in appendix 5E was used to obtain additional trial values for °n. (Appendix 5E also represents a first attempt to deal with qualitative materials closure for SRS.) In all cases the input stream was assumed to consist of lunar maria regolith, with values for In averaged from published data (Phinney et al., 1977) and listed in table 5.3. Following earlier work, for simplicity all efficiencies En were taken to be 093 (Rao et al., 1979; Williams et al.,1979).
The closures calculated from these data are plotted against extraction ratio in figure 5.23. (Data for the human body are included for purposes of comparison.) Note that 100% closure (C = 1) is achieved for the "U.S. Industrial" estimate (84 elements of the spaceprobe "seed") at R = 2984; for "Demandite" (28 elements) at R = 1631; and for the appendix 5E "LMF" (18 elements) at R = 45. This suggests that the fewer the number of different elements, and the more common and more ef