ETH Zurich researchers sought to bypass some of the limitations of evolutionary robotics by training a "mother robot" to autonomously assemble children robots out of component parts to see how well they move. The researchers note this removes the problem of having what works well in simulation not performing as well as expected in the real world.
Once built, the child robots' movements are observed and evaluated, and then the machines are disassembled and their constituent elements returned to the assembly line to be rebuilt into new robots. Successful software "elite" designs are carried over to the next generation without alteration, and they also are mutated or crossbred to produce the rest of the successive generation.
Five experiments were run, with 10 robot generations built, evolved, and improved in each instance. The elite designs did not always perform as well when they were retested, even though their designs were identical. Although the researchers attribute this inconsistency to significant variances in the behaviors of some of the robots, they observe "evolutionary pressure tends to select more consistent ones over generations, and usually repeatable genomes remain over generations."
Each experiment yielded a fitness increase of more than 40 percent over 10 generations, and the researchers envision this method to be complementary to simulations.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA