# New AI Finally Solved The Hardest Animation Problem!

## Метаданные

- **Канал:** Two Minute Papers
- **YouTube:** https://www.youtube.com/watch?v=nHBgc_oNfQw
- **Дата:** 31.08.2025
- **Длительность:** 5:14
- **Просмотры:** 66,494

## Описание

❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers

Guide:
Rent one of their GPUs with over 16GB of VRAM
Open a terminal
Just get Ollama with this command - https://ollama.com/download/linux
Then run ollama run gpt-oss:120b - https://ollama.com/library/gpt-oss:120b

📝 The paper is available here:
https://diffusecloc.github.io/website/

📝 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://twitter.com/twominutepapers
Thumbnail design: Felícia Zsolnai-Fehér - http://felicia.hu

## Содержание

### [0:00](https://www.youtube.com/watch?v=nHBgc_oNfQw) Segment 1 (00:00 - 05:00)

Here is an animation technique  that…oh no. What is going on here? This problem is a bit like an  orchestra with a genius conductor,   but the musicians are useless inflatable tube men. Okay, so we need a new AI animation technique that   we can control properly. Now I’m not  saying that animated movie and video   game characters can’t walk. Of course they  can - although…. not sure about this one. Now these classical animation techniques are  produced for every single motion by an artist.    However, AI-based techniques  offer something incredible:   just taking a soup of motion captured animations,   and learning how to use them and weave them  together. But there is a problem. The AI-based   animation techniques that are controllable,  often fail to create physically viable motion.    And there are ones that are realistic, but  those are unfortunately not easily controllable. So, does this new technique perform  the impossible? Let’s see the intro   sequence with the new method called  Diffuse-cloc…oh yes! Excellent! This is   the best of both worlds. Controllable,  realistic motion from a soup of data. And here’s the best part: it can  also do 5 other amazing things too! One, it can do static obstacle avoidance. It won’t  walk into walls. That’s a start, but now, check   this out. Dynamic avoidance too, where you put  together a bunch of AI people, and they hopefully   won’t bump into each other, when everyone  is moving. Yup, I think this one checks out. Two, it works for longer animation sequences too. Three, my favorite. Now hold on to your papers  Fellow Scholars for…generalization! This character   only saw jumping sequences on the ground, but now.   Three pillars. Let’s see if it can do it. One,   two, three! Excellent! I think this  really highlights how powerful these   learning methods are. They can do things in  ways that they haven’t seen before. Amazing. Four, you can specify two or more poses,  and it will generate the motion between   them. This is not possible with most  other diffusion-based AI techniques. Five, of course, you can grab this  character and try to disrupt its   movement. With previous methods,  this was…well, was it fun? Yes,   very fun. Was it disrupted? Oh yes! So how did the  new method do? Now as you see, it is not perfect,   it still looks a bit like someone who had a  couple drinks. But it is pretty resistant to   these perturbations. Loving this. This is  what we researchers do in our work hours,   and we call it work, and we even get paid  for this. Best job in the world. And this is   an incredible leap forward. So how is all this  possible? How did they do all this magic? Dear   Fellow Scholars, this is Two Minute Papers  with Dr. Károly Zsolnai-Fehér. Dr. Carroll. So remember, all this should work  from a bunch of unconnected input   motions. No information on how to weave  them together. to   use these in new situations. No obstacles, no  other people. This is all learned behavior. And this new AI is like teaching  a dancer not just the steps,   but also how to feel the rhythm a few  seconds ahead - so every move already   anticipates what comes next. The dancer isn’t  blindly following choreography anymore - it   can improvise while staying graceful.   That’s exactly how this system weaves   states and actions into one seamless motion.   Absolutely incredible. What a time to be alive! And all this is trained on a single GPU in  24 hours, something you can easily do in a   Lambda instance and then run it anywhere. And what  excites me most is that this is all zero-shot. No   retraining, no task-specific tuning - just one  model that learns to walk around walls, jump   over obstacles, play nice with others, and even  respond to a controller. Imagine game characters,   VR avatars, or robots that just move naturally  out of the box. And this is only the beginning.    Just imagine what we will be able to do  two more papers down the line. My goodness.

---
*Источник: https://ekstraktznaniy.ru/video/12151*