An Algorithm to Automatically Create and Optimize Robot Designs within Seconds
Designing a robot can be an incredibly time-consuming task, since one has to juggle a myriad of interconnected considerations that range from the robot’s physical form, to their sensory and motor systems and desired performance.
Generally, the process can take human designers months or even years of trial and error when designing, prototyping and testing new robots. While the recent advent of evolutionary algorithms to help automate the design process have helped, these methods are inefficient, and overly reliant on energy-intensive supercomputers and massive datasets, and still require some human-led tweaking along the way.
To tackle this problem, researchers at Northwestern University have developed an AI algorithm that is capable of automatically generating novel, optimized robot designs on a personal computer within seconds, without being hampered by a human designer’s preconceived notions about what may or may not work.
‘Instant Evolution’
Sam Kriegman, who is an assistant professor of computer science, mechanical engineering and chemical and biological engineering at Northwestern and the study’s lead author, commented on how remarkably fast the team’s approach was in a recent statement.
“[Our] AI-driven design algorithm… bypasses the traffic jams of evolution, without falling back on the bias of human designers,” said Kriegman, who is also credited with the 2020 development of self-replicating organic robots called xenobots. “We told the AI that we wanted a robot that could walk across land. Then we simply pressed a button and presto! It generated a blueprint for a robot in the blink of an eye that looks nothing like any animal that has ever walked the earth. I call this process ‘instant evolution.'”
With that simple request for a land-walking machine, the team’s algorithm then generated a never-before-seen robot in 26 seconds. Admittedly, it’s a bizarrely shaped thing, a molded and 3D-printed block of malleable silicone that looks riddled with holes.
But there is a rationale to its weird shape. As the team explains, starting with a small block of material about the size of a bar of soap, the AI was then able to iteratively assess the flaws of the robot’s physical shape and locomotive abilities, updating and ‘evolving’ it each time. At first, it could only jiggle, then on subsequent tries and with the help of some air injected into it, the AI’s robot could then bounce, shuffle and then finally walk — after only nine attempts.
Even more interesting is the robot’s final form — it has three legs, fins on its back and is punctured with holes — a structure that a human designer would most likely never come up with.
Kriegman noted that while the holes may look random, they in fact do help the prototypical bot walk much better than if it didn’t have the openings, because it makes it more lightweight and flexible: “We don’t really know what these holes do, but we know that they are important, because when we take them away, the robot either can’t walk anymore or can’t walk as well.”
Top left (A): graph comparing number of simulated prototyping attempts in previous studies (blue dots) to the team’s work (orange dot). (B) shows a photo of the mold used to create the team’s final prototype. (C) is the algorithm’s initial, randomly generated form, with red representing the bot’s ‘muscles’. (D-L) show the 9 design revisions that iteratively improve the robot’s walking ability.
In addition, Kriegman points out that human designers often tend to design bots to look like objects that are already familiar to them. In contrast, he and his team believe that AI can help to quickly generate new possibilities and ideas, outside of human limitations and the old and lengthy trial-and-error regime.
“Now anyone can watch evolution in action as AI generates better and better robot bodies in real-time. Evolving robots previously required weeks of trial and error on a supercomputer, and of course, before any animals could run, swim or fly around our world, there were billions upon billions of years of trial and error. This is because evolution has no foresight. It cannot see into the future to know if a specific mutation will be beneficial or catastrophic. We found a way to remove this blindfold, thereby compressing billions of years of evolution into an instant.”
The team believes that this near-instantaneous and fully automated design process will help to quickly address difficult problems that require customized solutions on-demand — from creating robots tailored to specific tasks like search and rescue after a disaster, repairing infrastructure, to personalized medical bots that are able to diagnose and treat disease more efficiently — all without having to wait months or years for a solution.