Human beings teach one another quite naturally. Other animals don’t teach, but they do learn from one another by copying behaviors that lead to positive results, like getting food. Machines can learn, in a very specialized sense of the word “learn,” but they never teach one another.
Until now. Google is teaching robots to teach.
The way robots learn things looks chaotic and random to human beings, but it’s essentially a matter of taking the information from multiple experiences and identifying the actions that get closest to the desired result. For one computer, that can mean a whole lot of inefficient time spent taking random actions. A Swiss study hurried the process along by adding evolution — copying the most successful artificial neuron pathways to create new generations of robots. It took hundreds of robot generations to complete the task.
But machines that were able to share their learning with one another, as humans do, could speed up the process considerably.
That’s where Google comes in.
By networking a group of robots and allowing them to share their experiments’ outcomes, they were able to produce much faster results on a task that is notoriously hard for robots: opening a door. They programmed robots to figure it out for themselves. Then they allowed the robots to share their learning. Finally, they had a human being teach the skill to a single robot, which shared the new skill through the cloud with other robots.
The last method accomplished the task more quickly than developing a program and programming the robots would have done.
Google may be the most important artificial intelligence company in the world, and the actions they’re taking will certainly make a difference for industrial motion control in the future. In the meantime, we can help you keep your Rexroth electric motion control system in trim. We have decades of specialized knowledge and the largest stock of factory refurbished exchange units in the U.S.