There are a number of very impressive examples of how artificial
intelligence can accomplish things that humans simply can’t, but nothing
is cooler than seeing AI married to robotics. OpenAI, the artificial
intelligence research group that counts Elon Musk among its founders,
has been working on honing the skills of its neural networks in various
ways, but this one-handed robot demonstration is probably its most
impressive feat yet.
The bot was tasked with solving a Rubik’s Cube with one human-style hand. This might not sound all that impressive, but it becomes a lot more interesting when you learn that the bot’s creators never actually told it how to accomplish its goal.
The pair of neural networks working in tandem to observe the cube and
control the hand were simply instructed to make all the sides of the
cube match. The AI then trained itself in a computer simulation,
learning how the movements of the hand affect the cube and how gravity
affects the cube as it’s being manipulated.
Multiple simulations can be carried out at the same time, and OpenAI
says that this method of training has allowed the AI to gain thousands
of years worth of experience in a very short period of time. Today, the
robot can quickly solve the cube under just about any condition.
As
the robot became more skilled at completing the task, the researchers
ramped up the difficulty by adding external factors. The team
experimented with a number of things, such as placing a rubber glove on
the hand to hinder its movement, physically fiddling with the hand as it
is trying to work, or even wrapping rubber bands around the bot’s
fingers. All of this helped the robot become more reliable at solving
the puzzle.
OpenAI says that this type of work is crucial in the
progress of multi-purpose robots that will require dexterity on par with
a human. Surely this will come in handy when the robots finally rise up
and crush us.
COMMENTS