Google's DeepMind project is teaching the computer about motor skills and the coordination required to walk.

Machine Learning

In the animation below, we can see the computer making mistakes with motor skills, recognizing the mistake, learning from that mistake, trying again, and eventually finding a way to achieve its goal.

Since the year 2000, we have seen robots such as Asimo setup to execute hand crafted or motion captured sets of commands from researchers in a very well structured environment.

While this is an amazing breakthrough in motor skills for a computer, the robot was not very capable on it's own. It did not seem to have the ability to learn from it's mistakes like we are seeing in DeepMind. Some of the latest improvements to Asimo do appear to suggest that the robot is learning more about it's environment, but it just does not seem to be learning quite the same.

Google's DeepMind was never shown how to walk, instead it trains itself by imitating the creation of motion capture data for standing up, walking, running, etc. After millions of attempts, lots of failures and lots of learning, the machine is now able to create new movements and it can transition between different types of movement all on it's own, with no direction from Google researchers.

Take a look at the video below from the research team at Google DeepMind to get a really nice explanation from the masterminds themselves!