Robots that learn from experience and can solve novel problems — just like humans — sound like science fiction. But a Japanese reasearcher is working on making them science fact, with machines that can teach themselves to perform tasks they have not been programmed to do, using objects they have never seen before.
In a world first, Osamu Hasegawa, associate professor at the Tokyo Insitute of Technology, has developed a system that allows robots to look around their environment and do research on the Internet, enabling them to “think” how best to solve a problem.
“Most existing robots are good at processing and performing the tasks they are pre-programmed to do, but they know little about the ‘real world’ where we humans live,” he told AFP. “So our project is an attempt to build a bridge between robots and that real world,” he said.
The Self-Organizing Incremental Neural Network, or “SOINN”, is an algorithm that allows robots to use their knowledge — what they already know — to infer how to complete tasks they have been told to do.
SOINN examines the environment to gather the data it needs to organise the information it has been given into a coherent set of instructions. Tell a SOINN-powered machine that it should, for example: “Serve water”. Without special programmes for water-serving, the robot works out the order of the actions required to complete the task. The SOINN machine asks for help when facing a task beyond its ability and crucially, stores the information it learns for use in a future task.