Industrial robot learns by trial and error

Fanuc is the world's largest manufacturer of industrial robots that use reinforced training to independently figure out how to accomplish their tasks.



In Tokyo, inside a modest-looking office building, lives an unusually intelligent industrial robot developed by the Japanese company Fanuc. Give him a task: take the widgets from one box and put them in another, and he will try to figure out how to do this all night. In the morning, the car will master this job, just as if it had been programmed by a specialist, reports technologyreview .





Fanuc demonstrated its product last December at the International Exhibition of Robots in Tokyo. Industrial robots are able to solve their tasks with the highest accuracy and speed, but usually they need to be programmed very carefully to teach, for example, to capture an object. It is difficult and time consuming, and means that such robots, as a rule, can only work according to a strictly defined algorithm.



A Fanuc robot uses a method known as “reinforcement training” to form its idea of ​​the task. He tries to capture objects with the manipulator and during this process fixes his work on video. Each time, regardless of the success of their actions, the machine captures an image of the object, which is later used to improve the algorithm of actions using “deep learning” or data processing in a neural network. Over the past few years, “deep learning” has proven its effectiveness in pattern recognition.

')

“After eight hours of training, the robot successfully performs 90 percent or more of the actions taken during the assignment, which is comparable to its programming by a specialist,” explains Shohei Hido, chief researcher at Preferred-Networks, a Tokyo-based machine learning company.



Experts in the field of robotics believe that reinforcement learning can simplify and speed up the programming of robots that are used in factories. Earlier this month, Google released details of its own research on using reinforcement learning, which teaches robots to capture objects.



Last August, Fanuc invested $ 7.3 million in Preferred-Networks. And in December, the companies demonstrated a self-learning robot at the International Exhibition in Tokyo.



One of the biggest potential benefits of this approach to learning is that the process can be accelerated if several robots work in parallel and then share the information they receive. Thus, eight robots working together for an hour can learn that one machine masters in eight hours. “Our project is focused on distributed learning,” says Hido. "Imagine hundreds of factory robots that communicate with each other."





Video showing the robot self-learning process



This form of distributed learning, sometimes referred to as “cloud robotics,” has great potential in both research and industry.



“Fanuc has a good market position for the development of this technology, as they supply robots for many factories around the world,” said Ken Goldberg, professor of robotics at the University of California at Berkeley. He adds that cloud robotics in the coming years is likely to change the modern concept of robots.



Nevertheless, he notes, the application of machine learning for robotics is a difficult task, since it is more difficult to control the behavior than, for example, to recognize objects in images. “Reinforcement training is a huge advancement in pattern recognition,” says Goldberg. “The problem with robotics is that people are multi-threaded. Accordingly, unlike robots, we are able to perform the necessary actions to solve a huge number of tasks at the same time. ”



Fanuc is not the only one who develops such robots. In 2014, ABB is a Swedish-Swiss company specializing in electrical engineering, power engineering and information technology. The company has invested in the project Vicarious . However, the fruits of these investments are still not visible.