By using Darwin’s principle of natural selection, researchers from Cornell University’s Creative Machines Lab got these virtual robots to evolve into proficient (albeit goofy) walking machines.

It was an experiment that led to some rather bizarre — and laughable — styles of locomotion.

The algorithm devised by the researchers was fairly straightforward. As biological evolution has shown time and time again, a simple set of rules, along with a ton of patience, can produce some rather remarkable things.


In this case, a research team led by Jeff Clune created an evolution simulator by endowing (relatively immobile) soft robots called soft-voxels with four basic building blocks to work with, namely muscle (shown in red), soft tissue support (teal), expanding and contracting muscles (green), and bone for hard support (blue). And importantly, they also programmed the system such that the faster bots would reproduce more. Speed, therefore, became a beneficial mutation (or adaptation) which served to increase a voxel's reproductive fitness.

Once these parameters were set, all Clune and his team had to do was press the start button and let evolution do the rest.


“Evolution did all the heavy lifting: there is no human in the loop after we start the Darwinian process," noted Clune in a YouTube comment. "It is definitely evidence for evolution doing impressive things, not for intelligent design.”

And indeed, after 1,000 generations, the system produced a series of fairly efficient walkers — some stranger than others. Perhaps unsurprisingly, bipedal locomotion emerged from the simulation, but so too did other techniques, like crawling, jiggling, flapping, and jumping. Some of the soft robots appeared to be animal-like, while others simply looked like crawling blobs of multi-colored pixels.

Writing in their paper, the researchers noted that locomotion performance increased as more materials were added to the voxels, and that “diversity and form and behavior can be increased with different cost functions without stifling performance, and that organisms can be evolved at different levels of resolution.”

Looking to the future, these findings suggest that scaled-up systems could be used to evolve “a large diversity of complex, natural, multi-material creatures” — creatures that could never be designed by straightforward engineering alone.

Read the entire paper: “Unshackling Evolution: Evolving Soft Robots with Multiple Materials and a Powerful Generative Encoding.” (pdf)