How To Train Your Robot: Bionic Elephant Trunk Learns Like A Baby
Precision control has always been something of a difficult hurdle in robotics - at least, for laymen. While it's certainly possible for a robotics expert to program a machine for precise tasks, an average user is more or less out of luck. Their choices are basically limited to either hiring on an expert.
That might not be the case for much longer, thanks to innovative new software developed by Jochen Steil and Matthias Rolf of a Bielefeld University in Germany. Using a process known as "goal babbling," the two researchers have developed a program that lets a robot to learn, through trial and error, how it needs to move to perform particular tasks. This is thought to mimic the way a baby learns to grab things by continual reaching.
The robot they're using was designed as a proof-of-concept by German engineering firm Festo back in 2010, to show that a trunk formed of 3D-printed segments could be controlled by an array of artificial pneumatic muscles. At the time it was developed, however, the trunk shipped without any sort of precision control software. It could perform rudimentary tasks - like grabbing a bottle or shaking someone's hand - but that was about it.
"They deliver it without much control. You can try, but the arm will be centimetres from where it should be, which is no good," explained Steil. The solution, he continued, was simple: design a robot with muscle memory.
The trunk - which looks more like one of Doctor Ocotpus's tentacles than anything taken off an elephant - remembers its position when changes are made to the pressure in its pneumatic tubes. This allows it to slowly create a map of the necessary pressure changes for a particular type of motion - in short, allowing it to "learn" how to move like a child would. Unlike a human infant, however, the trunk can also be'forced' into a number of different positions and eventually trained to adopt them on command - for example, replacing light bulbs or picking cherries.
Try to move the trunk into a different position, and it'll resist - initially - before eventually yielding and following the movement. At that point, the behavior is considered learned. If the tendril is moved to the same position again, it will go there easily and willingly.
It should be quite clear what such a technology represents. Although it's too rudimentary right now to have much use in the real world on its own, I imagine it would be fairly easy to prepare it for such. Robots equipped with muscle memory could be trained to carry out virtually any task with a minimal degree of technical knowledge. With the correct software and on-board, we could very well see machines step into a wide variety of roles previously thought impossible.
Not only that, with the development of this technique, robots have become that much more human. How long, I wonder, until we design a human-like machine capable of 'growing up' as we do? And what might the ramifications of such a thing be?
Those are questions we are in no way equipped to answer here, but it's interesting food for thought, no? In the future, consumers and professionals alike might train their robots in much the same way one would train their pets.
Like What You're Reading? Sign Up For My RSS Feed Here!
Robotics And Artificial Intelligence Blogger