Google DeepMind just taught itself to walk


#1

First AlphaGo, now this. Check out this humorous video.

Heck yeah. I’m going to walk like this from now on.


#2

AlphaGo has become a titan.


#3

That walking makes me want to see what it’d look like if it taught itself how to place stones.


#4

LOL, the human model reminds me of two things, namely …

  • of how AD(H)D kids often run, flapping their “wings”, and …
  • (of course) of Monty Python’s Ministry of Silly Walks

#5

It’s funny, but it’s also an absolutely amazing AI achievement.

I imagine if they put some cost to moving limbs, such as we humans incur, it would stop waving its arms around.

But getting AI to learn stuff for itself … general purpose real world stuff (albeit simulated) … that’s a big deal.


#6

Yeah, it would be interesting for them to model some sort of energy expenditure and then create some models that clear the course with the least amount of energy vs. least amount of time.


#7

This also shows that how we walk is relatively optimal when it comes to having two legs, considering that deepmind probably tested itself a thousand (probably much more, knowing computers)different ways before finally coming up with this. Evolution is awesome!
However I feel like the fact that it can clear walls higher then itself is somewhat unrealistic, maybe they could use a maximum amount of energy in a single step, as well as the least amount of energy thing that jgk and GreenAsJade suggested. Maybe then it would learn how to use its arms to hoist itself over walls, and maybe not flail so much.


#8

Have you seen parkour artists? They can get over pretty high walls, in a similar way that google does.