Researchers at Brandeis University in Waltham, Mass., have successfully programmed a computer to simulate the evolutionary process in a population of simple robots. The computer tested each successive generation of primitive robots to distinguish which could crawl the farthest. Then, the program itself refined the robots' design, replicating the process of adaptation intrinsic to evolution.
Hod Lipson and Jordan Pollack, the designers of the simulation program, call it a "robotic bootstrap." They also claim that the program, known as Genetically Organized Lifelike Electro Mechanics (GOLEM), marks the first time that automatically designed electromechanical systems have been manufactured robotically.
The program began with a population of 200 ma-chines. Initially, these de-vices were comprised of bars and actuators that served as building blocks of structure, along with artificial neurons acting as building blocks of control. GOLEM then conducted the evolution simulation. At each phase or generation, each machine's fitness was determined by its ability for locomotion. Such ability was measured by the net distance the machine's center of mass moved on an infinite plane in a fixed duration. In an iterative manner, the program selected the fittest machines and created more-evolved offspring by adding, modifying, and removing building blocks.
Within 300 to 600 generations, some robot species became extinct while others converged toward optimal designs. With the help of a 3D printer, the computer then built a handful of the fittest survivors from plastic. Using this 3D printer, or rapid prototyper, the program created a complete robot within its 8- by 8- by 12-in. parameters (see the figure). Human assistance was needed only for the attachment of motors and electronic "brains," which evolved alongside the physical parts in the simulation.
According to the researchers, the experiment's most surprising finding was the program's discovery of symmetry. Symmetry's importance wasn't specified within the program code. But in the course of the evolutionary process, the program itself found that balanced designs tended to crawl more efficiently. Since symmetric machines are more likely to move in a straight line, greater distance is covered. As a result, the program decided that symmetric machines were more desirable. Also, the robots' hardware and artificial-neuron software evolved concurrently, much like how the brain and body components of biological species have evolved jointly.
The GOLEM experiment and others of its type offer promise for the future design of self-replicating robots. Such devices would be invaluable to the exploration of extreme environments. NASA's interest in the notion of self-replicating robot factories in space missions reaches back to the 1960s. GOLEM's creators say that the evolutionary process must accept feedback from the live performance of its products if the process of self-replication is to be effective. For now, the GOLEM simulation is available in a screensaver version named Golem@Home.
For more information about the GOLEM program experiment, visit http://golem03.cs-i.brandeis.edu.
