Robots aren’t always so dextrous. They spend a lot of time falling over. Getting them to perform agility tasks that come naturally to humans can be hard to code, but they’re easy to demonstrate.
That’s why engineers at the AI Lab of SoftBank Robotics have been teaching robots to learn how to perform tasks demonstrated by a human, which they then practise over and over again until they get them right.
In the video below, you’ll see the company’s Pepper robot learning how to play that game where you have to catch a ball in a cup. To begin with, humans help the robot by guiding its arm.
Every success was logged, and every failure noted. After each failed attempt, the robot would alter its movement slightly to try and get closer to the solution. After 100 tries, it was able to catch the ball with a 100 percent success rate.
Softbank says that the task was accomplished using the free software library dmpbbo, which you can get here if you want to do a bit of robot-ball-catching yourself.
Eventually it’s hoped that the same principles could be used to teach robots other tasks that would otherwise require lengthy, complex programming and testing.