This AI Learns Faster Than Anything We’ve Seen!

This is Genesis and it’s a bit like a  universe simulation but with god mode,   where we build, break it down, and redo it better. And hold on to your papers right now Fellow  Scholars, because we’ve done nearly a 1,000   videos over 10 years, and you will see some  of the most beautiful simulations you ever saw   here. This is Genesis, a universal physics  engine available for all of you for free. You see, there are so many kinds of different  physics simulation systems out there,   and each of them can do one little thing  extremely well. Soft bodies, fluids,   granular simulations, particles, honey,  light simulations. One kind of experiment,   one system. If you want something else, you  have to implement a different algorithm. However, not here. Genesis puts all of these  together into one unified system and it took   my breath away. Not just because I  mean…look at this incredible beauty,   but there is more here. One, I also really  like about this work that it explains itself. And in a moment, I’ll tell you how this system  might change the world around us as we know it.    This technique might even cook your dinner  soon, and fold your laundry. Okay, so how? Well, step number one. Let’s generate  realistic simulations quickly. How quickly?    Can it be faster than MuJoCo, the library  that DeepMind bought for a fortune? Well,   NVIDIA’s Isaac Gym is much faster than that, it  cannot possibly beat that, right? Excuse me, what?    Are you seeing what I am seeing? Up to 80 times  faster, and its speed is in the…what? Hundreds of   millions of frames per second? Is this really  true? And what in the paper is that good for? Yes, it is true, but for a highly  optimized benchmark case, of course,   not in all cases by far, I’d like to  emphasize that. But still…wow. Here   is a thought experiment to demonstrate its  potential. At 244 million frames per second,   it can simulate so fast, that it goes from the  birth of Jesus Christ up to today, four times   over. And for every single hour of your life,  it can study for 30,000 hours of in-game time. Of course, it does not simulate the whole world,   but it can generate 30,000  different worlds in parallel. And now, we are ready for step number  two. What is that? Dear Fellow Scholars,   this is Two Minute Papers  with Dr. Károly Zsolnai-Fehér. Step two, let’s not just  look at these little worlds,   but also put a little robot in them to study them. Here is DeepMind’s soccer AI system that  started from zero, and then 3 years later,   it could do this. 3 ingame years later of  course, in our time, it’s just a few hours.    These little AIs started out like this, and when  they studied for 10 years, they ended up here. So now, imagine these little robots studying  for 30,000 years for every hour of our   lives. Or even better, let them study in  30,000 different worlds, one hour each.   Imagine this many different kinds of warehouses,  kitchens to work in, sandy beaches to walk on.    And by the time we upload the AI into a real  robot, it will be incredibly helpful. This is what   we call sim2real. I can’t even imagine how capable  it will be, because 3 and 10 years were amazing,   what could 30,000 years of learning  look like? What a time to be alive! And that is why this work might change  the world around us as we know it.    Teaching real robots to perform useful  tasks around us. And in the simulation,   we can have different kinds of robot morphologies  and have them learn how to be a good robot.    Including a robot that looks like a hand, or  this cute soft robot that worms its way into   this episode. Or of course, the super cute Disney  robot that steals everyone’s hearts. And then,   when they proved that they are useful and safe,  only then can they go out there in the real world. Now, we’re not done yet, genesis  has more goodies. It can generate   character animation too. These  look absolutely incredible. Wow.   It can generate interactive worlds  too, and objects separately within   these worlds. It is also a generative data  engine that can create new things from a   simple text prompt. I can’t wait to fire up  a GPU instance in Lambda and give it a try.

Especially because get this, it is designed to  be differentiable. Oh my! So, what does that   mean? What a differentiable programming system  does for us is that we can specify an end state,   which is the blue ball on the black dot, and  it is able to compute the required forces   and angles to make this happen. You can kind  of create desired futures with it. We talked   about this 500 papers ago. You can hear about  pretty much everything on Two Minute Papers. The user guide, code, library, everything is  available for you for free in the description.    Now, it is very difficult to get Youtube  to recommend this kind of content. However,   we tried it before, and found that if  I can get you to leave a kind comment,   it really helps. So I would like to ask every  single one of you to press like and leave a   kind comment so we can make videos about more  beautiful works like this. That would be so cool. Now, just to highlight both the good and  bad, we don’t shy away from that around   here. So two things. One, not every case is  as good as benchmark cases, so real-world   performance will be less, depending on what we  do. The 244 million frames per seconds was a   thought experiment with plenty of disclaimers.   That is always important to highlight. Also,   there is some criticism online about the claims  in this work, and there is a detailed answer to   the criticism as well, these are super useful  reads, both available the video description.