NVIDIA's New AI Broke My Brain

[00:00] Let’s see what is going on here. This is me  around 9am. A bit wobbly, steps are unsure,   yup, that checks out. Now then, give me my  fake badge. Thank you sir. Hehehe, no one  

[00:16] noticed. Now let’s proceed to the next step of my  mastermind plans. Let’s eat all their food. Wait,   they noticed. Proceed to the next  step. What was that? Oh yes, run!

[00:34] Now, jokes aside, look at that. Sign up  for this one baby. Oh yes, please mow   my lawn. That is excellent. Rake the leaves!  Perfect. Hey, don’t slack off, that’s my job!

[00:48] Okay, so what is going on here. Let’s start with  the good news, this is a new teleoperated robot   controller and more. They call it Sonic. Now  the work here is not the robot, but the software  

[01:06] controlling it. At least in this footage, watch  until the end and you might get surprised. This   means there is a human performing these movements,  and the robot is able to understand these motions,  

[01:20] and then translate them to a bunch of joint  positions in 3D space. It’s kind of insane   that this is possible. But it will just get  better and better as we continue the video.

[01:34] So, before you ask, yes it can do kung  fu. Provided that you can do kung fu. It   understands whole body movement, so you can get it  to crawl into some space you don’t want to go to.  

[01:49] And that is super useful, people are already  using robots for that. Why? Well, chiefly,   for exploring under explored and dangerous  areas. This means tons of useful applications,  

[02:04] for instance, a variant of this could  help save humans stuck under rubble,   or perhaps later, even explore other  planets without putting humans at risk. But that’s still nothing. Because this is a  multimodal system. Meaning that the input can be  

[02:22] almost anything. So, you say that I don’t have to  pretend to mow the lawn to actually mow the lawn,   because where is the fun in that? Well, just  tell it to do that. Can you? Well, currently,  

[02:40] for simpler tasks, like moving around or behaving  like a monkey, yes you can! Absolutely incredible. And I love how expressive it is.  You can ask it to walk happily,  

[02:53] And you know, just the fact that it is stable  and does not fall is remarkable. Previously,   even in simple characters in simulated worlds, you  needed thousands and thousands of tries to teach  

[03:10] them to just be able to walk without falling.  And now, this, is a huge leap forward. Wow. But it gets better, we said multimodal.  Yup, that means that the input can also  

[03:26] be music. I’ll show you the dancing, but  not the music because of Youtube reasons,  

[03:38] And we haven’t even talked about the most insane  part of the whole thing. Now hold on to your  

[03:50] papers Fellow Scholars, because this runs with  about 42 million parameters. That is a neural   network so simple, it can run so easily on your  phone it barely notices it. It may even run on  

[04:07] your toaster these days. That size is absolutely  nothing. This is an incredible achivement.

[04:19] Okay, but how? How is that even possible? Dear  Fellow Scholars, this is Two Minute Papers with   Dr. Károly Zsolnai-Fehér. Well, first, it  looked at 100 million frames of human motion  

[04:32] to understand what we do and how we do it. The  incredible thing is that this system does not   require human-made action labels, so we don’t have  to explain our movements. It just watches the raw  

[04:45] motions and figures out how to transition  between tasks without any unnatural pauses! So then, your multi-modal input goes in, a video  of you, your voice, music, or just text. A motion  

[04:58] generator turns these into human motion, and the  human encoder processes it into a latent space,   and then a quantizer converts it to universal  tokens. Once again, universal tokens, that is key,  

[05:13] you’ll see a bit later. Then, the decoder  translates these tokens into motor commands. But there is a big problem. Learning to convert  one to the other is super hard. First of all,  

[05:30] robots do not work like humans, that  is one of the fundamental challenges. So if the user commands you to turn around, it  should be turning around. Okay, sure. But how  

[05:43] fast exactly? You don’t want to try to turn 180  degrees too quickly, because you would fall apart. To solve this, in their research paper, they  propose what they call a root trajectory  

[05:57] spring model. This dampens sudden, quick user  commands so the robot does not get injured.   Yes, robots can get injured  too, which is kind of hilarious.

[06:10] Now there is an exponential term as a function of  time. What is that? That is a physical brake. As   time increases, this term rapidly shrinks to 0,  which forces the whole mathematical expression to  

[06:26] decay smoothly. This serves two goals: one,  the robot does not injure itself and two,   it will settle at a target position without  oscillating back and forth forever. Nice.

[06:41] Now, do the dampening too much, and  of course, you’ll get a little slug   that can’t get anything done, so it’s  really tough to do well. Well done folks. Now, all this took 128 GPUs and 3 days to  train. That is expensive. But here’s the key,  

[07:01] after the training is done, the final product  is so lightweight, we don’t need this kind of   hardware to run it at all. In fact, all  of the models showcased in these videos   will be given to all of us for free, forever.  They run on your phone, easy-peasy. That is  

[07:21] incredible. Open research for the benefit  of humanity. Love it, thank you so much. This project is led by professor Zhu and Jim Fan,   who I love dearly. Jim started the humanoid  robots lab at NVIDIA just 2 years ago,  

[07:38] and they are raining research papers on us,  breakthrough after breakthrough. Insanity. And to compress all this human movement  knowledge down into a tiny little AI  

[07:51] controller that can be used by any of  us is simply a stunning achievement. It turns out, training a good AI requires coding  good thinking into a machine. But, surprisingly,  

[08:04] we ourselves can also learn a lot of good  life advice from this kind of thinking too. For instance, the model compresses a messy,  diverse soup of inputs into a kind of pure,  

[08:17] abstract token. You know, in life,  when asking other people for advice,   you will inevitably hear everything,  and its opposite too. That is also a   big soup of inputs. But try to look at all of  them, side by side, and you’ll find that they  

[08:34] often share an underlying truth. This works,  as is showcased by this incredible project too. And note that this work is not the end of  anything, this is just a start. An early  

[08:48] work at a nascent area. Two more papers  down the line, and I really hope this is   going to start folding my laundry and cooking  my lunch. That would be amazing. What a time to  

[09:01] be alive! And this is not some proprietary  nonsense, this is open knowledge and open  

[09:26] just dropped. If you are interested in hearing  more hopefully soon, subscribe and hit the bell.