Watch NVIDIA’s AI Teach This Human To Run! 🏃‍♂️

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today is going to be a really fun day  where we do this, and this. In a previous episode, we talked about creating  virtual characters with a skeletal system,   adding more than 300 muscles and teaching  them to use these muscles to kick, jump,   move around and perform other realistic human  movements. You see the activated muscles with red. I am loving the idea, which, turns out,  comes with lots of really interesting   corollaries. For instance, this simulation  realistically portrays how increasing the   amount of weight to be lifted changes  what muscles are being trained during   a workout. These agents also learned to  jump really high and you can see a drastic   difference between the movement required  for a mediocre jump and an amazing one. And now, scientists at NVIDIA had a crazy idea.   They said, what if we take a similar model, and   ask it to learn to walk from scratch.   Now that is indeed a crazy idea,   because they proposed a model that is  driven by over 150 muscle-tendon units,   and is thus very difficult to  control. So, let’s see how it went. First, it started to…umm…hello? Well, A+ for  effort, but unfortunately, this is not a great   start. But don’t despair! A little later, it  learned to…well, fall in a different direction,   however, at least some learning is hopefully  happening. Look. I wouldn’t say that it has   finally taken the first step, but at least it  is attempting to take a first step. Is that good   news? Let’s see! Oh yes, yes it is! Because a  little later, it learned to jog. This concept   really works! And, if we wait for a bit longer,  we see that it learned to run as well. Fantastic! Now, let’s have a closer look and see if the  colors of the muscles indeed show us which   ones are activated at a given moment. And that’s  right! When slowing the footage down, we see the   difficulty of the problem - and that is, different  tendons need to be moved every single moment. So, while we look at this technique learning  other tasks, we ask one of the most important   questions here, and that is, how fast did it learn  to run? It had to control a 150 different tendons   continuously over time, without falling. So how  long did it take? Did it take days? And now,   hold on to your papers, because it hasn’t taken  days. It only takes minutes. After starting out   like this, by the 17-minute mark, it has learned  so much that it could jog. How amazing is that? And that is one of the key value propositions of  this paper. It can not only teach this AI agent   difficult tasks, but it can also learn up to  15 to 17-times faster than previous techniques.    That is absolutely amazing. Bravo! So, it seems  that we now have learning-based algorithms that   could teach even a complex, muscle-actuated  robot to walk. What a time to be alive! Thanks for watching and for your generous  support, and I'll see you next time!

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today is going to be a really fun day  where we do this, and this. In a previous episode, we talked about creating  virtual characters with a skeletal system,   adding more than 300 muscles and teaching  them to use these muscles to kick, jump,   move around and perform other realistic human  movements. You see the activated muscles with red. I am loving the idea, which, turns out,  comes with lots of really interesting   corollaries. For instance, this simulation  realistically portrays how increasing the   amount of weight to be lifted changes  what muscles are being trained during   a workout. These agents also learned to  jump really high and you can see a drastic   difference between the movement required  for a mediocre jump and an amazing one. And now, scientists at NVIDIA had a crazy idea.   They said, what if we take a similar model, and   ask it to learn to walk from scratch.   Now that is indeed a crazy idea,   because they proposed a model that is  driven by over 150 muscle-tendon units,   and is thus very difficult to  control. So, let’s see how it went. First, it started to…umm…hello? Well, A+ for  effort, but unfortunately, this is not a great   start. But don’t despair! A little later, it  learned to…well, fall in a different direction,   however, at least some learning is hopefully  happening. Look. I wouldn’t say that it has   finally taken the first step, but at least it  is attempting to take a first step. Is that good   news? Let’s see! Oh yes, yes it is! Because a  little later, it learned to jog. This concept   really works! And, if we wait for a bit longer,  we see that it learned to run as well. Fantastic! Now, let’s have a closer look and see if the  colors of the muscles indeed show us which   ones are activated at a given moment. And that’s  right! When slowing the footage down, we see the   difficulty of the problem - and that is, different  tendons need to be moved every single moment. So, while we look at this technique learning  other tasks, we ask one of the most important   questions here, and that is, how fast did it learn  to run? It had to control a 150 different tendons   continuously over time, without falling. So how  long did it take? Did it take days? And now,   hold on to your papers, because it hasn’t taken  days. It only takes minutes. After starting out   like this, by the 17-minute mark, it has learned  so much that it could jog. How amazing is that? And that is one of the key value propositions of  this paper. It can not only teach this AI agent   difficult tasks, but it can also learn up to  15 to 17-times faster than previous techniques.    That is absolutely amazing. Bravo! So, it seems  that we now have learning-based algorithms that   could teach even a complex, muscle-actuated  robot to walk. What a time to be alive! Thanks for watching and for your generous  support, and I'll see you next time!