Building machines from a multitude of parts on a production line was a triumph of the industrial revolution.

A lathe can make many useful things, but it can't duplicate its parts and put them together to make a baby lathe identical to itself. To do so would require deep knowledge of its own construction-the lathe would have to know itself. No machine ever has, but recently engineers and scientists at NASA have shown that a self- replicating machine could be built. With a relatively modest investment, a self-  replicating factory could be built by early in this century. The Japanese have reached a similar conclusion.

Self-replicating machines would produce factories as well as products, so that self- replicating machines would be the ultimate tools for increasing productivity. One application investigated by the NASA group was to use a self- reproducing machine on the surface of the moon to make solar cells, which could be cheaply hauled to vast solar- power satellites that would beam solar energy back to earth. The researchers considered the task of making one million solar cells out of the moon dust. A $1 billion, 100-ton conventional factory would take about 6,000 years to perform this task. A self- replicating factory of the same initial size and cost could do the job in just 20 years, because it makes not only Solar cells but more solar-cell factories. Marvin Minsky, artificial- intelligence expert at the Massachusetts Institute of Technology and former president of the American Association for Artificial Intelligence, recently wrote, "Teaching computers how to build copies of themselves could begin a flood of automatic self-replication machines making more machines at very low cost. Then we'll have to learn to cope with the resulting explosive growth of wealth and productivity."

It was not until 1948 that scientists became convinced that machines could be taught to replicate themselves. In that year John von Neumann, the Hungarian- American mathematician who helped design the first stored-programmeme computer, gave a series of historic lectures at the University of Illinois. He showed that to duplicate itself, a device must have available its own blueprint, some manipulators- hands, essentially- and a set of rules for building things. The beauty of von Neumann's approach was that he was able to prove, with mathematical rigour, the possibility of building self-reproducing machines.

The simplest self-reproducing machine von Neumann imagined would sit in a giant stockroom filled with replacement parts- extra arms, legs, eyes, circuit boards, and the other paraphernalia from which it was built. The machine would have a memory tape containing all the instructions needed to build a copy of itself from these spare parts. Using its robot arm and its ability to move around, the machine would find and connect parts. The tape programmeme would instruct the device to reach out and pick up a part, look to see if it's the right one, and if not, put it back and grab another. Eventually the correct one would be found, then the next, and the two joined in accordance with the master checklist. The machine would continue to follow the instructions, never knowing what it is making, until it finishes assembling a physical duplicate of itself. But the new robot would be "uneducated"; it would have no instructions on its tape. The parent would then copy its own memory tape onto the blank tape of its offspring. The last instruction of the parent's tape would be to activate its progeny. 

Some simple examples of machines that reproduce without human intervention do exist. Small mechanical contrivances capable of self-assembly from simpler parts are easy to build. Roger Penrose, a British physicist, fabricated a set of simple blocks that could hook together using a set of ratchets and levers. When single parts are placed in a box-a track shaped like a long trough- and shaken, they do not join together. However, when an interlocked, two-block unit is placed in the box and shaken, a simple form of reproduction takes place. Collisions between the two-  block unit and other parts in the box cause new two- block units to form, each identical to the original. Penrose successfully tested replicating units of up to four blocks, using several different block types.

It is important to recognize that a complex robot is not essential to make a useful machine. It is possible to design a modular growing robot system, like the early lathes that could produce parts that could make better lathes, which could then produce parts that made better lathes, and so on. But the exercise has not been done.

I