Clever clumsiness: A self-taught walking robot

A group of researchers at UC Berkeley (including EE Prof. Sergey Levine, grad student Tuomas Haarnoja and undergraduate researcher Aurick Zhou) and Google Brain have used maximum-entropy reinforcement learning to make a quadrupedal robot teach itself to walk.   It taught iself through trial and error in a mere two hours before researchers introduced the machine to new environments, like inclines and obstacles, where it adapted with ease.