Researchers in the US have trained AI to design a robot from scratch

Fed with a basic prompt, the AI system was able to develop a blueprint for a rudimentary walking robot within seconds.


With the public launch of ChatGPT last year, artificial intelligence (AI) and its impressive capabilities exploded into the mainstream.

Now, researchers from Northwestern University in the US have built on that momentum by developing an AI system capable of designing functional robots in mere seconds.

The study, published in the Proceedings of the National Academy of Sciences journal, demonstrated AI generation of novel robot designs without the influence of human biases or previous blueprints for the first time.

The system was fed a basic prompt: design a robot capable of walking on a flat surface. The AI algorithm was then able to generate its own blueprint for a walking robot.

The blueprint takes the form of a 3D-printed mould which is then filled with liquid silicone. The result is a spongey, somewhat formless object with the internal design helping it to mimic a walking motion.

“It actually starts from scratch. And the way that it works is it generates this random body that kind of looks like a sponge and it evaluates its behaviour,” explained Northwestern’s Sam Kriegman, who led the work.

“Thanks to a little neat mathematical trick, it can kind of see into the future and tell how changes to the body will affect its behaviour, make it better or worse.”

The squishy, misshapen robot may not look particularly refined or impressive but it belies a significant leap forward in AI-designed tools, according to Kriegman.

While traditional approaches to AI often involve colossal datasets and energy-intensive supercomputers, one of the attributes of this new AI system is that it can function on a simple personal computer.

The team found that seemingly random holes in the robot’s body played a vital role, enhancing flexibility and weight distribution, further highlighting the algorithm’s ability to innovate beyond conventional human designs.

Kriegman described it as ‘evolution in action’, as the AI-driven design algorithm bypassed the traffic jams of evolution, without falling back on the bias of human designers.

“And I’m hoping that this generates new ideas and new hypotheses and leads eventually to new useful technologies,” he said.

“As we give the AI more things to play with, we hope that it can result in more complex and more useful robots.”

For more on this story, watch the video in the media player above.

Video editor • Aisling Ní Chúláin

Source link

2023-10-16 09:27:10

#Researchers #trained #design #robot #scratch

Learn More →