ITU professor in Prestigious Journal: Game Technology Can improve AI-Development
Professor Sebastian Risi publishes paper in Nature Machine Intelligence on how to improve training of robots with popular video game technology.
Digital Design DepartmentResearchartificial intelligencecomputer gamesSebastian Risi
Written 6 August, 2020 09:27 by Simone EtwilMeyland
Teaching robots to solve new tasks can be easier if researchers and companies use Procedural Content Generation (PCG) – a technology developed for video games. This is the main point in a paper by Sebastian Risi, professor at IT University of Copenhagen and co-founder of Modl.AI. The article was published Monday in the prestigious scientific journal Nature Machine Intelligence.
“We hope that PCG will become a more mainstream tool in the work with artificial intelligence and robotics. PCG can save a lot of money and will hopefully enable robots to solve more diverse and complex tasks,” Sebastian Risi explains. He wrote the paper with former ITU-researcher Julian Togelius who is now an associate professor at New York University and another Modl.AI co-founder.
From Video Games to Robotics
Game developers have been using PCG for years. Basically, the technology enables the computer to generate content itself, so a developer does not have to. By generating levels, maps, characters etc. the games become replayable.
According to the paper by Sebastian Risi and Julian Togelius AI-researchers are beginning to use PCG as well. But instead of using the technology to create game content, they use it to produce training data for robots.
Training data is essential in the development of artificial intelligence. If you want a robot to solve a task you will have to teach it in a computer simulation. But there is a gap between simulation and reality, and many robots will not work in real life without additional training, which is a very costly process.
But with PCG the computer can generate more training data and reduce the need for real-life training. A company called Open AI taught a robot to solve a rubrics cube with PCG, and because the computer generated different textures, sizes, shapes, frictions etc. it was easier for the robot to solve the cube in real life as well.
“With PCG we can speed up the development of AI and robots and lower the costs. On a longer term we can possibly create a more general and robust artificial intelligence. Evolution taught humans to handle constantly changing challenges, so we’re very good at adapting. This ability we want to pass on to the robots, so they can solve new challenges they have not met before.” Sebastian Risi says.