Virtual beings develop bodies that help them learn

The virtual creature swung its four arms like tentacles, propelling itself forward. It crept up the hill then rushed to the other side. It looks like “octopuses walk on land,” says Agrim Gupta. This strange creature developed a body of its own. It also learned its own method of moving. This mix of evolution and learning could help engineers build new types of robots, Gupta said.

A PhD student studying computer vision at Stanford University in California, Gupta is like a grandfather to this octopus-like creature and hundreds of other odd-looking virtual creatures. He created the ancestors that gave rise to these creatures. He called them unimals, meaning “universal beasts.” The term reflects the fact that they can evolve into so many different body forms. Some resemble real animals. Others are quite strange.

The team found that the animal’s body type affects its ability to learn new things. We tend to think of learning as something that happens in the brain. But, Gupta notes, “your body plays a huge role in what you can learn.” The type of world you live in also matters.

If robots can evolve in simulations, they might develop their own forms that work better, think Gupta and his colleagues. Then the engineers were able to build bodies they never dreamed of for themselves.

So they gave it a try. Unimal who learns to move in a more complicated simulated world ends up with a body more suited to learning. Gupta and his group describe this in Nature Communication end of October.

“I love this job,” said Sam Kriegman. He was not involved in the research but knows a lot about the topic. He works on evolutionary robotics at the Wyss Institute. It’s part of Harvard University in Boston, Mass. He also worked at the Allen Discovery Center of Tufts University in Medford, Mass. Robotic engineers tend to imitate the bodies they see in nature. That’s why many robots resemble real animals, such as dogs or humans.

The inspiration for designing unimals came from animals, says Agrim Gupta. He thought they might have evolved to look and move like real animals. In fact, they didn’t look the way he expected. “Not even one,” he said.

go round and round

Animal species evolve with small, random changes in their genes. Changes that provide new advantages make it easier to survive. Computer scientists can now emulate this process in code. This is how the Gupta team did it.

To begin with, they give their unimal bodies very much like stick figure beasts. Each has one round head. Straight segments sticking out of this head. They branch into other segments, forming body parts that resemble arms, legs or tentacles.

Machine learning includes deep learning and neural networks

More than 500 randomly generated unimals are thrown into a virtual world, which is very similar to a video game. In the simplest games, each unimal must traverse a flat landscape. It’s looking for ways to move around using computer machine learning models. Machine learning is a type of artificial intelligence (AI) that allows computers to practice skills until they master them.

In this case, the machine learning model controls the unimal body. At first, when the model knows nothing about moving, the body will twitch while trying random movements. If one movement brings the unimal closer to its goal across the landscape, the model learns to repeat that movement. The further the unimal crosses the landscape, the higher the score in the game.

The bouncing starfish

Then, the unimals were divided into four groups. Whichever group member has the highest score can evolve. Let’s imagine the winner looks like a starfish. When it evolves, its body changes randomly. For example, he may lose some of his legs. Or, all of its legs might grow a new segment. Or one can be longer and the other shorter. In this latter case, the limbs become lighter. Then “starfish can bounce more easily,” explains Gupta.

Then, all the unimals from the original group of four returned to the flat virtual world along with the new starfish. They remember nothing from their first trip around the world. They all have to start from scratch, go round and round until something works. Again, they all score and face off in groups of four to see who will progress next.

This process is repeated, over and over. Every time a new unimal is created, the oldest one dies. If he does a good job, then he will evolve several times before dying. That means he leaves behind a group of children and grandchildren who could have done better. Over generations, unimals have gotten better and better at traversing landscapes. They don’t remember anything from past experiences. That’s because the point is not to traverse the landscape. This is to develop a better body in learning to move.

Educators and Parents, Cheat Sheet List

Weekly updates to help you use Science News for Students in the learning environment

Thank you for signing up!

There was a problem registering you.

Facing challenges

The flat world is just the beginning. Gupta and team went through the same process again with a new random group of unimals in a undulating landscape. And in the third world, the unimals have to push a cube across multiple targets across a bumpy landscape. This is very difficult to master. However, by combining learning and evolution, unimals emerged that could handle it. One evolved two limbs like hands which were used to propel the cube.

The team then tested all the unimals in the new type world. It has obstacles that have never been encountered before. They had to climb up and down steep slopes. They have to push the ball towards the target (which is much more difficult than the cube because it can roll easily). Again, the unimals remember nothing of what they have learned. All they had was a body shape that worked well in one of the three original worlds.

Unimals bodies have to follow some rules. They must have symmetrical right and left sides. Also, they cannot have more than 10 limbs, and each limb cannot branch more than twice.

Unimal that has evolved in a third world — a world of mounds and cubes — “learns new tasks better and faster too,” notes Gupta. Why? Their bodies have adapted to help them solve different kinds of problems.

For example, unimals with hands can use them to propel the ball. Unimals from the flat world don’t need hands, so have a much harder time controlling the ball. Having the right body, Gupta points out, “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 have to use computers to help us be more creative and come up with new types of robotic bodies,” Kriegman said.

It won’t be easy to turn a simulated creature into a reality, he added. The real world is much more messy and complex than the simulation. A body that works well on a computer may not work well in real life. However, Kriegman says, “this problem can be solved.”

#Virtual #beings #develop #bodies #learn