Robots amaze humans with their amazing performance. They collect cars, search collapsed buildings for survivors, and serve restaurants. But often a simple task is very difficult. A closed door, for example, is a difficult obstacle for robots. Japanese engineers have now taught a robot how to open a door. They write in the journal that their learning method also works with other daily activities robotics science†
Modern robots have often found their way in a way deep learning is called. Whether playing chess or riding a bike, the robot is told what to do and masters the technique by trying and learning from its mistakes. The Google computers that crushed the world chess champions learned the game with this deep learning.
Expensive and time consuming approach
The Japanese write that it is an expensive and time-consuming approach. Robots are often good at the task they have been taught, but feel awkward when conditions change. For example, if the robot finally knows how to open the door, it can no longer leave the building because the door swings the other way.
The Japanese argue that it is best to allow the robot to take advantage of both worlds. Not only through deep learning, but also through legacy instructions. That’s why they didn’t let their robot detect it from the closed door – by the way, they break the task down into six pieces, including operating the latch or keeping the door open. They also taught him hundreds of lessons in opening doors.
The two-sided approach has paid off. After a few hours of training, the door no longer had any secrets for the robot. He easily passed it on 97 percent of attempts. And when he was asked to go back and forth constantly, he kept working for half an hour, without any wrong step. “Our experience shows that this dual-learning approach allows robots to operate in a real world with unexpected challenges.”
Two robots put an Ikea chair together
The art of robotics was taken to unprecedented levels by engineers in Singapore