A virtual creature swings four tentacle-like arms and pushes itself forward. It crawls up a hill and rushes down the other side. It looks like “an octopus walking on land,” says Agrim Gupta. This strange creature developed its own body. It also learned its own way of moving. This mix of evolution and learning can help engineers building new kinds of robots, says Gupta.
Gupta, a PhD student studying computer vision at Stanford University in California, is sort of a grandfather to this octopus-like creature and hundreds of other strange-looking virtual critters. He created the ancestors that gave rise to these creatures. He calls them unimals, which stands for ‘universal animals’. That term reflects the fact that they can evolve in so many different body shapes. Some resemble real animals. Others are quite bizarre.
The team found that an animal’s body type influences its ability to learn new things. We tend to see learning as something that happens in the brain. But, Gupta notes, “your body plays a huge role in what you can learn.” The kind of world you live in also matters.
If robots could evolve in a simulation, they could develop their own shapes that work even better, Gupta and his colleagues thought. Then engineers could build bodies they would never have imagined on their own.
So they tried it out. Unimals learning to move in more complicated simulated worlds ended up with bodies better suited to learning. Gupta and his group described this in: nature communication last October.
“I was excited about this work,” says Sam Kriegman. He was not involved in this research, but knows a lot about the subject. He working on evolutionary robotics at the Wyss Institute. It is part of Harvard University in Boston, Massachusetts. He also works at Tufts University’s Allen Discovery Center in Medford, Massachusetts. Robot engineers tend to copy bodies they see in nature. Therefore, many robots resemble real animals, such as dogs or humans.
An animal species evolves with small, random changes in its genes† Those changes that give it new benefits make it easier to survive. Computer scientists can now mimic this process in code. This is how Gupta’s team did it.
For starters, they gave their unimals bodies that look a lot like animal stick figures. Each has a single round head. Straight segments protrude from this head. They branch into other segments, forming body parts that resemble arms, legs, or tentacles.
Just over 500 randomly generated animals are thrown into a virtual world, much like a video game. In the simplest game, each unimal must cross a flat landscape. It finds out how to move using a computer model of machine learning. machine learning is kind of artificial intelligence (AI) that allows computers to practice a skill until they master it.
In this case, the machine learning model controls the animal’s body. In the beginning, when the model knows nothing about movement, the body swings around while trying random movements. If one movement brings the animal closer to its target to traverse the landscape, the model learns to repeat that movement. The further the unimal moves through the landscape, the higher the score in the game.
A bouncing starfish
Later, the unimals are split into groups of four. The member of the group with the highest score may evolve. Let’s imagine the winner looks a bit like a starfish. When it evolves, its body changes in a random way. For example, it may lose part of its legs. Or all its legs can form a new segment. Or one gets longer and the other shorter. In the latter case, the limbs become lighter. Then “the starfish can jump around more easily,” explains Gupta.
Later, all the animals from the original group of four go back into the flat virtual world along with the new starfish. They don’t remember anything about their first trip around the world. They all have to start over, roam around until something works. Again, they all get a score and compete in groups of four to see who gets to evolve next.
This process repeats itself over and over. Whenever a new unimal is created, the oldest one dies. If it did a good job, it will have evolved a few times before it died. That means it left behind a bunch of kids and grandkids who might do even better. Over many generations, animals are getting better and better at traversing the landscape. They remember nothing of past experiences. That’s because it’s not about crossing the landscape. It is to develop bodies that can learn to move better.
Take on challenges
The flat world was just the beginning. Gupta and the team went through the same process again with new groups of random unimals in a bumpy landscape. And in a third world, the unimals had to push a cube toward a target in a bumpy landscape. This was especially difficult to master. Combining learning and evolution, however, created unimagines that could handle it. One developed two hand-like limbs that it used to push the cube.
The team then put all unimals to the test in new types of worlds. These had obstacles that no one had encountered before. They had to go up and down steep slopes. They had to push a ball towards a target (which is a lot trickier than a cube because it can easily roll away). Again, the animals did not remember anything they had learned. All they had were body shapes that had worked well in one of the original three worlds.
Unimals that had evolved in the third world — the one with the bumps and the cube — “learned new tasks better and much faster too,” Gupta notes. Why? Their bodies were modified to help them solve different kinds of problems.
For example, the unimal with hands could use those to push a ball. Flat world Unimals didn’t need hands, so had a much harder time getting the ball under control. Having the right body, Gupta showed, “can greatly simplify the problem of learning a task.”
Engineers can’t always imagine the best body type for a particular robot. By combining evolution and learning, designers can generate and test thousands of new options. “We should be using computers to help us be more creative and come up with new kinds of robotic bodies,” Kriegman says.
It won’t be easy to move simulated creatures to reality, he adds. The real world is much messier and more complex than a simulation. A body that works well in a computer may not work so well in real life. Kriegman says, however, “these problems are solvable.”
#Virtual #critters #develop #bodies #learn