PR2 - Robot Drinking-buddy Training
Explore the world and combine - these are the two pillars on which human civilization is based. Well and greed, of course. Design cups alone were invented in the 21st century more than great inventions over the past century. Although, it would seem that it could be easier than a faceted glass by Mukhina and a cup holder.
However, if for us the process of mastering the subject world (and our own body) was relatively easy in the distant, distant childhood, then for robots learning how to cope with numerous
Therefore, the development of deep learning algorithms has remained the main direction in the development of robotics for the eighteenth year in a row. And where else to test such algorithms as not in the field with a wide variety of rattles: cogs, latches, keyholes and buttons.
The most capable child prodigy today is the brainchild of Willow Garage under the name PR2 (Personal Robot 2) . The guys somehow did not grow together with the first model, but the heirs are already actively moonlighting in hotels .
PR2 has several advantages . Naturally, these are high-quality manipulators akin to YUMI , good sensors, and most importantly, open source code. Usually, seven nannies have a child without an eye, but this time as nurses, there are specialists from the best universities in the world, so training moves by leaps and bounds.
In my head, PR2 has a bunch of fresh libraries, hardware drivers and workpieces that help scientists do something new and not re-teach the robot how to walk. PR2already knows how to open beer, go for coffee (doors, elevator, bring-serve), play billiards and - most importantly - hold objects. Such successes, by the way, became possible largely due to tactile feedback, so the robot will not flatten a jar of hot tea in front of your nose, but carefully put it on your table.
However, all this was back in 2013. In just a year and a half, the algorithms have grown to the point where PR2 can now brew your coffee. Attention: make coffee on an unknown coffee machine, following the instructions printed on a piece of paper. For example: “Hold a cup of espresso under a tap with hot water.”
The idea is that the robot independently navigates in a room full of objects, recognizes the necessary buttons and pens in a form without prompts and a pre-compiled knowledge base. All that is in his hands is a piece of paper with directions and objects.
Applause to the Cornwell developers!
Another literally recently published achievement in the education of robots by the deep learning algorithm "trial and error." This time, the program was written by scientists from the Berkeley Research Center.
In this case, the developers were able to reproduce the learning process, which is as close to human as possible. Naturally, little PR2 still has a lot to learn, but scientists have marked the road correctly.
Now, to set a task for a developer, it takes about 10 minutes of time. The machine takes about 3 hours to complete the task. This, of course, is a bit much (considering that in the real world it’s painful to step on the rake), however, this algorithm is just the beginning and scientists hope that the robot will begin to develop exponentially.