Jeff Clune, a visiting scientist, and his team in Hod Lipson’s Creative Machines Lab have made progress towards that futuristic world through mechanical simulations that utilize natural selection in robots.
Close your eyes and imagine a world where engineers could grow robots that learned and evolved on their own. Jeff Clune, a visiting scientist, and his team in Hod Lipson’s Creative Machines Lab have made progress towards that futuristic world through mechanical simulations that utilize natural selection in robots.
Jeff Clune, mechanical and aerospace engineering, said that he gained extra interest in science after learning about evolutionary encoding.“I’m interested in trying to evolve artificially intelligent robots. I want to know how natural selection on earth produced all the wonderful complexity that you see — like jaguars, hawks, and the human mind.”
Evolution is often difficult to study because it is a phenomenon that takes place over a long time frame and it requires a lot data to produce useful results. Whereas these potential studies may take many years to conduct, simulations such as those Clune works with can run in a matter of minutes.
Using evolutionary encoding, Clune had created a “grown brain” and input it into a simulated table robot. The robot is then asked to learn to develop a gait. At first, the table robots were not able to achieve lateral movement. However, using distance as a measure of fitness, Clune managed to evolve a natural gait out of an initially incoherent four-legged table. The final generation that was evolved moved much like a horse would gallop, with three of their four legs in synchronization.
Unlike traditional robotic programming, this natural selection process generates more efficient and natural gaits according to Clune. Also, there is less work involved because in cases where the programmer would be required to write out an entire gait, basic brain evolution achieves a better result with less work.
Along with Jason Yosinski, a Ph.D. student, Prof. Hod Lipson, mechanical and aerospace engineering, and Suchan Lee ’13, mechanical engineering, Clune took the next step and incorporated natural selection into a physical, 3-D robot which they printed using the 3-D printer in the Creative Machines Lab. To create the robot, they first generated it on a computer then sent it to a printer, which then crafted the robot layer by layer using a type of hard plastic. Using the same idea as the simulated table robot, the printed robot started out with a basic brain and eventually evolved a creepy, spidery gait.
“No human designed this—evolution did,” Clune said.
In addition to evolving movement in robots, Clune is also interested in evolving the physical shapes of the robot.
“All of those gaits were with the body being fixed, but we also wanted to evolve the body and see things flying and crawling,” he said.
Clune then ran simulations based on dynamic bodies as well as the evolving brains to create counterintuitive gaits. The results were synchronized, but also modeled natural movement. Some simulations scooted across the screen much like a walrus would, and some hopped like a rabbit or kangaroo. By evolving the body and mind simultaneously, Clune creates models which exhibit characteristics that traditional engineering could not achieve.
Clune along with Yosinski, and Lipson, have created EndlessForms.com, a website that allows any one with time to spare the chance to create and share new designs for things evolved through natural selection.
“The website really helped people understand evolution. They witness each step of the journey and can look up the lineage of every object online. Many universities and elementary schools have used this to try to teach evolution.”
Upon visiting the site, the user can choose to either evolve designs already manipulated by other users or start with a panel of randomly generated variations of a block. The website has produced designs such as butterflies and complex chess pieces — all of which would have been difficult to engineer without a substantial background in 3-D programming. However, through the natural selection simulation process, even ten year olds can generate stunning 3-D renderings of natural objects.
“When you put a human in charge of programming, you’re limited by the creativity and knowledge of the programmer,” Clune said. “However, when you put evolution in charge, the sky is the limit. This is going to change how engineering will work,” he said.