# NVIDIA’s New AI Cheated At Parkour…And Got Fixed!

## Метаданные

- **Канал:** Two Minute Papers
- **YouTube:** https://www.youtube.com/watch?v=AVeQfSab9to
- **Дата:** 08.07.2025
- **Длительность:** 6:35
- **Просмотры:** 61,207
- **Источник:** https://ekstraktznaniy.ru/video/12267

## Описание

❤️ Check out DeepInfra and run DeepSeek or many other AI projects: https://deepinfra.com/papers

📝 The paper "PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers" is available here:
https://michaelx.io/parc/index.html

📝 My paper on simulations that look almost like reality is available for free here:
https://rdcu.be/cWPfD 

Or this is the orig. Nature Physics link with clickable citations:
https://www.nature.com/articles/s41567-022-01788-5

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Sven Pfiffner, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi
If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers

My research: https://cg.tuwien.ac.at/~zsolnai/
X/Twitter: https://

## Транскрипт

### Segment 1 (00:00 - 05:00) []

This is an AI player trying to survive  in a challenging game environment,   ouch…and it ended up a bit like a parkour  simulator. It is also cheating like crazy,   but don’t worry, we won’t let it do that. You see, to teach it, we start from motion  capture data. You know, motions copied from   real humans. It’s a start…but it’s not much.   Only 14 minutes of data, you can’t really do   anything meaningful with that. Except if…well,  here is an incredible idea from scientists at   NVIDIA and Simon Fraser University in three steps.   One. Let’s use this data that we have. This is the   character that you see here. Now look at these  characters. What is going on here? Well, two,   they create new, randomly generated levels, and  three. They use a physics-based engine to create   new motions from the new levels and the data we  have. So why are two characters here? Why not   just one? And what is all this useful for? Well,  you’ll see that the result will be absolute magic. Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. You see, new, purely kinematic  motions here are dreamed up by the AI,   but these can have floating, foot  sliding, and other kinds of cheating,   so they have to be corrected by a physics engine  to create physically plausible motion. Then,   we add these new movements to our  initially small dataset and grow it. Then,   the cycle starts again. We’ll see in a moment  what this is good for, but wait a second. We generated new levels, but how do the  motions actually happen? What do we,   the character do on these levels? Well,  they generate paths within these levels,   and the character is supposed to follow them,  which can involve, climbing, jumping and more. So now the moment of truth! How did it do?   Well, after the first cycle of enriching the   dataset? Not too great. So, keep going,  one more round of enriching the dataset,   and then. Hey! Come on man, this  is cheating. What just happened? Well, note that the green characters  from earlier go through a path,   but not necessarily in a physically correct  manner. We need to correct their motion. And   after doing that, let’s see…now we’re  talking. 3 iterations of enriching the   small starting dataset with physics-based  correction, and now we are in business! And with this, the most incredible thing  happened. So what did it learn? Did it   just learn to do stuff that it had already  seen before? Nope. Not in the slightest. Check this out. One, it learned to combine  multiple motions together. Like jumping,   holding on to…well, not the papers, but  the cliff edge, and climbing up. And now,   hold on to your papers Fellow  Scholars, and look. Did it survive?    Oh yes, and it has crazy skills now…and to  think that it learned this by itself. Wow. And now comes my favorite. Testing whether this  is really intelligent, how do we do that? Well,   we test it on new environments it hasn’t seen  yet. And at this point, we understand that the   green character is just a dream, and the blue  is the one that with its physics corrected. And it seems like these two jolly chaps can finish  any level we can make them. So cool! Of course,   our choice is the blue one. So, they  can do lots of jumping and climbing,   I’ll be honest, at this point, I am  not too surprised by that. But now,   give me something truly different from the  previous levels. Climbing a monument, okay,   I love that. And there is one point that  I was really impressed with. Don’t blink,   because it is really easy to miss. Did you  see that? Look. After doing the first jump,   like most video game characters, it did not  stop to execute a second jump, but instead   it hopped forward on one leg. It didn’t even  stop, and it looked really natural. Loving it! But what we want now is more levels! During  the climbing of the spiral, I am liking this,   but all this running around made our  AI friend a little…. constipated. It’s   alright little AI, I won’t tell anyone about it. And we are Scholars here, we love to look  into the research works, and two things   really took my breath away when reading  the paper. One, every clip in the original,   small motion capture dataset was converted into 50  different terrain variations. That is incredible.    So a single recording suddenly became a  rich playground of environments. And two,

### Segment 2 (05:00 - 06:00) [5:00]

it does not require a huge cluster of  GPUs to train. Just one graphics card,   although an A6000, so an expensive one,  and it takes a while too, up to a month   to train. Still, the possibility of  doing it on one card is incredible. Now, limitations. Not even this technique  is perfect. The motion generation itself   is a bit slow. It can take about  25 seconds to create 1 second of   character movement on a GPU. That's like  waiting 4 minutes to see 10 seconds of   parkour! But just imagine what we will be  capable of two more papers down the line. So, teaching an AI to survive in the dire  world of the spire, even when constipated,   is now possible. How cool is that?   Hopefully coming to games and virtual   worlds near you soon. What a time to  be alive! It is absolutely crazy that   no one that I know is talking about this  amazing work. Only on Two Minute Papers.