Dear Fellow Scholars, this is Two Minute Papers
with Károly Zsolnai-Fehér. About 350 episodes ago in this series, in
episode number 8, we talked about an amazing paper in which researchers built virtual characters
with a bunch of muscles and joints, and through the power of machine learning, taught them
to actuate them just the right way so that they would learn to walk. Well, some of them
anyway. Later, we’ve seen much more advanced variants where we could even teach them to
lift weights, jump really high, or even observe how their movements would change after they
undergo surgery. This paper is a huge step forward in this
area, and if you look at the title, it says that it proposes multiplicative composition
policies to control these characters. What this means is that these complex actions are
broken down into a sum of elementary movements. Intuitively you can imagine something similar
when you see a child use small, simple lego pieces to build a huge, breathtaking spaceship.
That sounds great, but what does this do for us? Well, the ability to properly combine these
lego pieces is where the learning part of the technique shines, and you can see on the
right that these individual lego pieces are as amusing as useless if they’re not combined
with others. To assemble efficient combinations that are actually useful, the characters are
first required to learn to perform reference motions using combinations of these lego pieces.
Here, on the right, the blue bars show which of these these lego pieces are used and when
in the current movement pattern. Now, we’ve heard enough of these legos,
what is this whole compositional thing good for? Well, a key advantage of using these is that
they are simple enough so that they can be transferred and reused for other types of
movement. As you see here, this footage demonstrates how we can teach a biped, or even a T-Rex
to carry and stack boxes or how to dribble, or, how to score a goal. Amusingly, according
to the paper, it seems that this T-Rex weighs only 55 kilograms or 121 pounds. An adorable
baby T-Rex, if you will. As a result of this transferability property, when we assemble
a new agent or wish to teach an already existing character some new moves, we don’t have
to train them from scratch as they already have access to these lego pieces. I love seeing
all these new papers in the intersection of computer graphics and machine learning. This is a similar topic to what I am working
on as a full-time research scientist at the Technical University of Vienna, and in these
projects, we train plenty of neural networks, which requires a lot of computational resources.
Sometimes when we have to spend time maintaining the machines running these networks, buying
new hardware or troubleshooting software issues, I wish we could use Linode. Linode is the
world’s largest independent cloud hosting and computing provider, and they have GPU
instances that are tailor-made for AI, scientific computing and computer graphics projects. If you feel inspired by these works and you
wish to run your experiments or deploy your already existing works through a simple and
reliable hosting service, make sure to join over 800,000 other happy customers and choose
Linode. To reserve your GPU instance and receive a $20 free credit, visit linode.com/papers or
click the link in the description and use the promo code “papers20” during signup.
Give it a try today! Our thanks to Linode for supporting the series and helping us make
better videos for you. Thanks for watching and for your generous
support, and I’ll see you next time!