Use your hand to grab the handle, which lies on your desk – a very simple task, but the robot is still a very difficult task. Therefore, to teach robots to understand the rise of casual items, we created the Google research team, which has helped 14 robots to cope with the task.
The standard way to solve this problem would be the examination of the robot environment, bc, followed by drawing up a plan of how to grab an object, and then pick it up. However, in the real world a lot of things can change between the development of the plan and its implementation.
Google is now using CNN (technology that complies with the latest requirements in the field of machine learning) to help their robots to predict the outcome of their actions on the basis of the input command signal and the camera. In other words, it is the interaction of hand-eye coordination robot.
The team says it took about 3,000 hours of practice and 800,000 attempts before they saw “rudiments of intelligent reactive behavior.”
“The robot observes its own grip and adjust their movements in real time. It also has a preliminary understanding of the behavior, how to isolate one object from a group“- Says the team. “All of these problems have arisen, of course, in training, and have not been programmed into the system.”
Researchers at Google say that the average failure rate without training were 34% for the first 30 attempts. After training, this figure dropped to 18%. Of course, this is not ideal, but the next time, when the robot will be run for you in an attempt to grab, remember, they are now 80% chance of success.