A baby deer can stand up for ten minutes after the time of his birth, and, for a period of seven hours to do so. In between those two milestones, he has shown a remarkable and frantic Fidgeting through the day, in order to find the right technique.
It is the idea of Artificial intelligence (AI), fed to the robotic. Autonomous robots, such as self-driving cars are now a familiar concept, but the Autonomous robots are not learning from it yet. Algorithms for reinforcement learning robot to learn in fact, by trial-and-error movements, but also the intensive human intervention in order to be used Each Time the robot falls, or is in addition to the training, to go out, you have to send it to him to cancel it, or return it to the upright Position.
Researchers at Google, and now it has become an important first step in the direction of the robot, which can learn from it, without any help, it is to navigate, how to communicate in a new role as a coach. Within a couple of hours, they managed to get a four-legged robot in a completely independent way, both in front and behind and to the left or turn to the right. To this end, the researchers were the only changes to existing modern Algorithms.
The work is based on a survey from about a year ago, the group met for the first time, as the robot could, in a real-world learning experiences. Reinforcement learning is usually carried out in a Simulation, a virtual DoppelgangeR for the robot to move through a virtual replica of the real-life environment to which the algorithm is robust enough to in real life. It is then transferred to the physical robot.
The robot avoids the damage to
With this method, it is possible to prevent, during the process of trial-and-error, damage to the robot and the environment, but they are in need of an easy-to-simulated environment. As the sand is in place, it is a robot and put his foot on it, or how deep of a mattress this sinks in, it is not so difficult to calculate that in the simulations, to be of service to you here.
Against this backdrop, the Google researcher decided to dispense with the simulation, and to train in the real world. They were able to develop an algorithm that is more efficient, with less experience, and therefore, the less the error that it requires, and within a couple of hours, you have your together. Because of the environment that have had changes, he was able to respond quickly to similar environments, with slopes, obstacles, and stairs.
A man had to take care of the robots, and hundreds of thousands of times, to step in, ” says Jie Tan, one of the authors of the article, and is a senior member of the robotics team at Google in the Brain. “I wouldn’t have thought of in the beginning,” he said.
This is a new Problem for the Team. First of all, it is excluded from the area in which the robot was allowed to explore, and it’s guaranteed that he / she performed various maneuvers at the same time. When the robot reaches the edge of the area, as he had learned how to get in front of him, he changed his direction, and he learned to go back to it.
The limits of the movement
In a second step, the researcher limits set forth for an experiment on the movements of the robot, and it did so in a careful, repeated falls is no great harm. After a Swim, they felt more ready for the algorithm, which in turn helped the robot to stand up.
With these optimizations, the robot to walk on different surfaces stand-alone, even on solid ground, with a Memory-foam mattress, and a ribbed doormat or she has learned. The present study shows the potential for new applications in the future you have to navigate on a robot, without human assistance, through the difficult and unknown Terrain.
“I think this work is very exciting,” said Chelsea Morris, an Assistant Professor at Stanford University, which co-operates with the Google, but not for the specific project involved. “The people, the process of making it is really, really hard. If the robot can self-directed learning program, you will learn more in the real world in which we live, rather than in a lab.”
Be adapted to other robots
Today, the concept does, however, require a System for the detection of the movement of the robot to determine its location, on the borders of Finland. It would not be in the real-world, viable.
In further steps, Google, the researchers want to adapt the algorithm to different types of robots, and to be able to do this, many robots that can learn and at the same time in the same room. The Problem with continuing to resolve the motion, ” says Tan, it is essential to make robots more useful.
“A lot of the posts are made by people, and you have the legs,” he said. “If a robot can use its legs, it can be in the human world to sail.”