Out of the digital ooze, robotic life
Designs that evolve without human aid

By Charles W. Petit

Maybe Jordan Pollack, computer scientist and robot maker, should call his
workplace at Brandeis University near Boston something other than a
laboratory. Nest or hive seems closer to the mark. In it, he and mechanical
engineer Hod Lipson run the Golem Project, a colony of machines that evolve
and give birth to other machines without human guidance. Last week, in the
journal Nature, the pair described their work as "Automatic design and
manufacture of robotic life-forms." Life-forms? That is something that
Lieutenant Commander Data might say on the Starship Enterprise: "Captain, a
scan of the planet reveals robotic life-forms!"

So far the life-forms–bundles of white tubes resembling crazily linked
sausages that lurch, slither, and hunch their way across lab bench
tops–aren't very advanced; their only talent is movement. But neither were
the first creatures to appear in Earth's primordial ooze. Massachusetts
Institute of Technology artificial-intelligence expert Rodney Brooks calls
these machine-bred robots a step toward the "ultimate dream of self-evolving
machines." They conjure up a tomorrow in which machines breed among
themselves, whelping devices whose abilities surpass human invention–a
vision at once hopeful and unnerving.

To get machines to design other machines, Pollack and Lipson borrowed a
strategy from nature: Darwinian evolution, played out not in a warm pond but
in the software of a computer. The computer was programmed with a set of
designs that were no more than disordered collections of struts, ball
joints, and electric motors, plus electronic circuit parts for a nervous
system. It randomly altered, or mutated, the initially useless designs.
Next, the computer chose the "fittest" mutants–those that showed hints of
locomotion–while killing off the others, in a digital version of natural
selection. It further mutated these chosen few and then repeated the process
over hundreds of generations, gradually evolving more capable robots.

The designs that eventually emerged might never have sprung from the human
mind–but they do move. Some work by hunching forward like inchworms, others
drag themselves along by walking on elbowlike protrusions, and still others
creep sideways, crablike. In some cases, the evolutionary process stumbled
onto the symmetrical body form common in living things, which makes
movement in a straight line simpler.

Putting natural selection to work in a computer isn't new. Half a century
ago, the pioneering Hungarian-American mathematician John von Neumann
imagined self-reproducing computer programs called "cellular automata."
Since then, artificial-life programmers have watched digital organisms
evolve as they compete for survival in software ecosystems. But until now,
the offspring have lived in cyberuniverses akin to video games. Even when
engineers used such methods to produce novel designs for real objects–some
jet engines, for instance, have features that evolved in computers–their
actual manufacture was under human control.

Newborn robots. The Golem Project, however, automatically hatches the
winners into the real world. An off-the-shelf, industrial "rapid
prototyping" machine builds the kitten-size robots from plastic, following
the evolved instructions rather than a blueprint penned by a flesh-and-blood
engineer. Watching the embryonic devices spontaneously emerge and begin
creeping about (after people click electric motors into place where the
computer tells them to) "is a little scary," concedes Pollack.

The whole process takes just hours or days, compared with the million-year
time scale of evolution in nature. But other researchers say the
machine-bred robots are a case study in the power of natural selection as a
design tool in the lab and in the natural world. "It is clear that evolution
can do things that people cannot possibly do on their own," says David
Fogel, chief scientist at Natural Selection Inc., a company in La Jolla,
Calif., that evolves solutions to problems in business and other fields. He
even thinks the Golem Project and other evolution-based technology could
persuade people who are skeptical about evolution to accept Darwin's ideas.
"The more people accept evolution as a design tool, the easier it is to see
that it has been used by nature."

Whether or not the embryonic robots are plausible apostles for natural
selection, they are likely to get still more lifelike before the project is
over. So far, the evolution stops as soon as the robots step from the
software into the real world. But Stanford University consulting professor
John Koza, a breeder of computer programs, would like to see the actual
plastic robots evolve generation by gene- ration into more adept forms. John
Holland, a University of Michigan pioneer in an artificial-evolution method
called genetic algorithms, wants the project to add cybersex, allowing robot
parents to pool their best features in their offspring, rather than using
mutation alone for improvement.

Inevitably, the work resonates with fears of robots running amok. Most
notably, Sun Microsystems chief scientist Bill Joy warned early this year
that a triple whammy of genetics, nanotechnology, and robotics could unleash
a "gray goo" of tiny, voracious, self-reproducing microbots marauding over
the world–a fear also stoked by recent work showing that robots can
collaborate in antlike swarms (box).

Pollack dismisses such fears. "We're far, far from that," he says. "[These]
are small plastic gizmos that look like toys." Yet the project's name plays
on the unease such work stirs. A golem, in ancient Jewish legend, is a lump
of clay sculpted to human form and animated through cabalistic magic. Some
variants are tireless servants that work without pay. Others make mischief
and must be killed to protect the community–a centuries-old foreshadowing of
Joy's warning.

© U.S.News & World Report Inc. All rights reserved.

Links to Pollack and Lipson:

You can see the influence of Neural Networks.

Download The Golem Project: "Live Truss" evolution program.  It's FREE.