Robots From Scratch?

Dr. Andrew L. Blais onlymice at javanet.com
Wed Aug 30 20:53:22 PDT 2000


I thought that lfs-discuss subscribers might find this interesting....

        Aping Biology, Computer Guides
        Automated Evolution of a Robot

        By KENNETH CHANG

          F or the first time,
           computer scientists
        have created a robot that
        designs and builds other
        robots, almost entirely
        without human help.

        In the short run, this
        advance could lead to a
        new industry of
        inexpensive robots
        customized for specific
        tasks. In the long run --
        decades at least -- robots
        may one day be truly
        regarded as "artificial
        life," able to reproduce
        and evolve, building
        improved versions of
        themselves.

        Such durable, adaptive
        robots, astronomers have
        suggested, could someday
        be sent into space to
        explore the galaxy or
        search for other life.

        But the quest to create artificial life also revives concerns
        that computer scientists could eventually create a robotic
        species that would supplant biological life, including
        humans.

        "Some things we probably can do we shouldn't do," said Bill
        Joy, chief scientist at Sun Microsystems, who wrote a
        recent article warning of the power of emerging
        technologies. "Just like we can kill things with DDT, but we
        shouldn't."

        For now, the robotic manufacturing system -- a computer
        hooked up to a machine that builds plastic models -- in the
        laboratory of Dr. Jordan B. Pollack and Dr. Hod Lipson at
        Brandeis University in Waltham, Mass., cannot create
        anything nearly as complicated as itself. Instead, it produces
        eight-inch-long contraptions of plastic bars and ball joints.

        When a motor and microchip are added, the automatons
        have one, and only one, ability: to crawl slowly. The fastest
        can scuttle along at a few inches a second.

        "They look like toys," Dr. Pollack, a professor of computer
        science, said. But, he added, "They were not engineered by
        humans, and they were not manufactured by humans."

        Dr. Pollack and Dr.

        Lipson, a research scientist, report their results in today's
        issue of the journal Nature.

        "This is the first example of pretty much 100 percent
        automated evolution of a machine," said Dr. Philip
        Husbands, a professor of artificial intelligence at the
        University of Sussex in England. "It's a rather primitive
        example, but it's the first step to something that could be
        quite significant."

        In the future, the technique could be used to design robots
        that assemble parts in factories, clean up chemical spills or
        vacuum a home.

        Because computers cost much less than human engineers,
        "It opens for the first time a more economical approach to
        robotics," Dr. Pollack said. "We can now essentially design
        for free and build for a few thousand dollars."

        The cost of designing a robot today typically runs from
        hundreds of thousands of dollars to millions of dollars, Dr.
        Pollack said.

        The computer in the Brandeis system had no idea what a
        successful design might look like. Instead, it was merely
        given a list of possible parts it could work with, the physical
        laws of gravity and friction, the goal of moving on a
        horizontal surface and a group of 200 randomly
        constructed, nonworking designs.

        Mimicking biological evolution, the computer added,
        subtracted and changed pieces in the designs. At the same
        time, the computer similarly mutated the programming
        instructions for controlling the robot's movements. After
        each step, the computer ran simulations to test the designs,
        keeping the ones that moved well and discarding the
        failures.

        After 300 to 600 generations of evolution and fine-tuning,
        the computer sent the design to a prototyping machine,
        used by manufacturers to build test models of product
        designs, to build the robot. Then, in the step that required
        human help, the researchers installed the robot's motor
        and microchip and downloaded the robot's programming
        instructions.

        Changing the initial configuration of the robot parts
        produced a different design and a different approach to
        locomotion. One pushes itself along. "It's kind of like an
        accordion," Dr. Pollack said.

        Another one "walks something like a crab," Dr. Lipson
        said. "It looks like it's crawling on the floor. It's quite
        surprising the diversity of solutions we get."

        In earlier work, other researchers used similar
        evolution-inspired algorithms to evolve imaginary
        creatures that existed only in virtual computer worlds or to
        design the programming instructions.

        The robots' evolution is currently a dead-end, as the
        designing computer never learns how well its designs work
        in the real world. The simple robots also have no ability to
        improve their performance. According to the researchers,
        the robots currently have the brainpower of bacteria. "We
        hope to get up to insect level within a couple of years," Dr.
        Pollack said. "There's no danger of Commander Data
        walking out of our fabricator anytime soon," he said,
        referring to the android character in "Star Trek."

        In future research, the Brandeis researchers intend to add
        sensors to the robots and improve the design programs.

        Future robots may also be able to exchange information
        among each other and learn from each other's experiences.

        As computer chips speed up and fabrication machines
        become more sophisticated, the robotic designers will
        produce robots that are more and more complex. Some
        have wondered what will happen when a robot can design
        and build something as complex as itself.

        For example, Dr. Seth Shostak, an astronomer at the SETI
        Institute in Mountain View, Calif., has suggested that
        researchers listening for radio signals from alien
        civilizations are more likely to first come across intelligent
        machines created by aliens.

        In an article in the April issue of Wired, Mr. Joy argued
        that scientists should perhaps deliberately steer themselves
        away from research that would create self-replicating,
        evolving, autonomous robots.

        With forethought, he said, computer scientists should be
        able to tap into most of the benefits of the emerging
        technology while avoiding the dangers.

        "This doesn't have enough of the pieces to be by itself
        dangerous," Mr. Joy said about the Brandeis work. But, he
        added, "We're on the road to somewhere where there's big
        issues down the road."

        Others working in the field are not as worried, even if
        technological advances make such devices possible. Dr.
        Ralph C. Merkle of the nanotechnology firm Zyvex and an
        adviser to the Foresight Institute, said that high costs would
        probably prevent the design of dangerous robots. Rather,
        robots would continue to be designed for specific tasks with
        little or no ability to evolve and adapt.

        "It looks like having a device to work at all is hard," Dr.
        Merkle said. "There is no desire to add additional
        complexity. Those systems do not look like they would be
        dangerous."

        The Brandeis researchers find the speculation premature.
        "Really, it's so far removed from anything dangerous," Dr.
        Lipson said about their work. "There are many other things
        to worry about before this."



              Copyright 2000 The New York Times Company








More information about the lfs-dev mailing list