When the robots are cooperating in order to learn

deep-reinforcement-learning-for-robotic-manipulationEveryone knows the paradox of the monkey, a few thousands of monkeys, placed in front of a machine to write, eventually type the entire of Hamlet by the force of the accident. The robot, thanks to Machine Learning may also eventually learn how to open a door by dint of trial/error.

For several years, researchers have shown that a robot could learn from itself a simple gesture. However, even a stain also site to open a door by turning a handle, it takes hours, or even days. Google researchers imagined connect multiple robots to power a “collective” intelligence and above all learn faster.

Several robots connceted to learn faster


Sergey Levine, of the Google Brain Team, Timothy Lillicrap of DeepMind, and finally Mrinal Kalakrishnan, robotician experienced at Google have interconnected neural networks of 4 robots have a simple problem: turn a doorknob and open it. As expected, this co-operation between robots led logically to accelerate the learning of the group phase. The researchers did vary the angle of the door, and thus allows the robots to raise the quality of their learning. At the end of the experience, researchers believe that the quality of this learning is higher than that as a single robot could achieve alone.

The researchers then compared their robots has a another trial, that of push various everyday objects arranged on a table in a receptacle. This event is no longer based on the trial and error but screws to highlight the quality of perception of the environment of the robot. There again, cooperation between robots has paid off. Researchers consider the benchmark for image recognition ImageNet with more than 1.5 million images would require several years of apprenticeship for a single robot, although in fact was cooperate more, learning can be achieved in just a few weeks. In addition to speed, learning several to obtain better quality learning. The tasks are still simple, but with the improvement of the algorithm, but also an increase in the number of robots connected to each other, they can develop skills well higher than those they have demonstrated so far.

deep-reinforcement-learning-for-roboti-manipulation-2 

 

Sources :
“How Robots Can Acquire New Skills from Their Shared Experience”, Google Research Blog, October 3, 2016
“Deep Reinforcement Learning for Robotic Manipulation”, Cornell University Library, October 3, 2016

Commentaires Facebook
Twitter Facebook Plusone Pinterest Linkedin
This entry was posted in Robotique and tagged , . Bookmark the permalink.