Robots learn 'sockification'

Robots developed at UC Berkeley shed light on how models can learn to complete complex tasks with 'perception mechanisms'.
Written by Charlie Osborne, Contributing Writer

Research in robotics have progressed at a rapid pace in the last few years, and yet, you cannot necessarily envision them performing delicate tasks like handling and folding up your laundry.

A team at the University of California, Berkeley, had other ideas. Can a robot perform a previously unseen task merely by watching, such as manipulating socks?

According to the engineering team's research project (.pdf), the aim was to create robots that are able to perform tasks by mimicking human actions, or learn through a series of pictures.

Pieter Abbeel, an assistant professor at the university, develops robots that are able to handle the steps required in performing a task correctly -- without individual programming for each separate stage. Some of his previous creations have learnt to perform acrobatics, tie surgical sutures and complete other delicate tasks.

Socks were chosen as they 'trace no easily-recognizable silhouette [..] and lend themselves to highly complex configurations: the socks may be rightside-out, inside-out, or arbitrarily bunched'. Therefore, the challenge issued to these PR2 robots is actually rather complicated:

  • The robot is presented with two socks.
  • It must classify them as either 'inside' or 'outside' and flip accordingly.
  • Once both socks are matched, it then has to pair them.

What separates these robots is the fact they are given 'perception mechanisms' that allow them to analyze what they experience, and then mimic various behaviors, rather than simply be programmed to handle designated tasks.

Abbeel writes that loosely structured environments cause issues for robotics, and 'poses a number of challenges to robot perception and manipulation'. Robotics is currently more suitable for fixed roles -- for example, within the manufacturing industry. However, when placed in an environment which allows a degree of freedom and requires a different set of skills, robots aren't necessarily developed enough to be useful.

In the video, we can see exactly how complex a task like folding and manipulating a piece of cloth actually is. Things we consider simple are actually quite a complicated series of events, and require manual dexterity and analytic thinking to complete.


This post was originally published on Smartplanet.com

Editorial standards