Computer scientists at Johns Hopkins University have shown how to teach robots through the use of positive reinforcement, an approach familiar to anyone using a reward to change a dog’s behavior, and the robot, called Spot, was able to learn in days instead of a month.
Lead author Andrew Hunt explained, “I have dogs so I know the rewards are beneficial and that was an inspiration for how to design the learning algorithm.”
He added: Unlike humans and animals who are born with very intuitive brains, computers are blank boards and they have to learn everything from scratch. But real learning is often achieved by trial and error, and the team achieved this by devising a reward system that treated robots like dogs. Where a dog might get a biscuit to do a good job, and robots get a numerical score.
The researchers envision this method could help train home robots to wash clothes and dishes – tasks that could be common on the open market and help elderly people live independently. “We don’t currently know how to program tasks like this – the world is so complex,” Heiger said. But work like this shows us that there is promise in the idea that robots can learn how to accomplish such realistic tasks in a safe and effective manner. ”