NVIDIA’s New AI Trained For 10 Years! But How? 🤺

📝 The paper "ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters" is available here:

📝 Our material synthesis paper with the latent space is available here:

❤️ Watch these videos in early access on our Patreon page or join us here on YouTube:

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Geronimo Moralez, Gordon Child, Ivo Galic, Jace O'Brien, Jack Lukic, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.

Chapters:
0:00 10 Years of training?
0:42 After 1 week
1:23 After 4 months
1:31 After 2 years
1:55 After 10 years!
2:25 How did they train for 10 years?
3:01 1. Latent spaces
3:52 2. Robust recovery
4:35 3. The controls are 👌
5:01 4. Adversaries
5:57 A great life lesson
6:15 The Third Law of Papers

Károly Zsolnai-Fehér's links:

[Mouxeware]

I'm a little upset they didnt have them fight each other. I would have loved a large scale fight.

[johnclark]

Like Ray Kurtzweil recently said, “In the next 10 years, we will see 175 years”. This isn’t in relation to AI progress, we’re just going to shove 175 years into 10.

[franticsledder]

I see them clipping the swords through their skulls sometimes. If the researchers added damage zones to the heads and other parts, I bet the AI would learn to be more careful how they swung those weapons.

[DiffYT]

Imagine using this system in a gladiator simulation game where the gladiators are randomly generated with different movement and combat "styles" with a skill level that changes how smooth and refined the gladiator is. Seeing millions of unique potential combinations battle it out in various arenas with various environments in duels and team battles, different weapons and an ai that makes use of the environment well could be a next level video game for sure.

[Atrazeal]

So basically you just need to put this trained AI into the Atlus Boston Dynamics robot and the matrix is no longer a work of fiction. damn.

[williamhanna]

Player animations has always been the most immersion-breaking part of 3rd person games for me, and I’m really hopeful we will see massive improvements in big games soon!

[OperationDarkside]

I wish, they'd put a kind of scanning or vision element into the network, so it can evade the boxes from the researchers and after 5 more years, we'll see generated matrix style fight scenes.

[BevansDesign]

It'd be interesting to see them model things like realistic reaction times and muscle strength/fatigue, so the characters don't have perfect recoveries when they get knocked over.

[w.d.gaster]

I'd love to see learning enemies in a mmo type videogame so both players and the game would adapt to have a constant challenge

[Andimax]

This would be amazing if you could put physically simulated characters like this into VR games like Bonelab. They would make some of the most incredible and dynamic enemies in video games

[konignickerchen]

I notice it is using flat terrain with the same variables every time, I wonder how well it generalized for different terrain, gravity and character proportions.

[mackenziestorey]

The way the sword works with latent space is really interesting. Can't wait to see where this research ends up after a few more papers!

[tizianonatali]

Looking forward to see them in non-levelled terrain

[bobdemtrich]

This would be really cool for open world games. Imagine your character starts with a set basic movements, through playing it develops it’s own unique style of combat based of your movements and style. Mixed with multiplayer to create a truly unique combat style, where everyone’s movement and combat is different and optimized to their play style.

[battycrease]

I think it would make an interesting game where the ai enemies learn how the player moves and how the player fights and throughout the game they learn to imitate and counter you as the game progresses and the final boss is basically like fighting against yourself.

[johnsonYT]

5:50

The from scratch simulation has so much unused potential.

Imagine if you used an animation like that for Halo's Flood parasite combatants.

So creepy. Just needs the right model, and environment.

[trevorhook]

I'd love to see the AI trained on reactions to hostiles with the characters having their own "fields of view" and reacting realistically to what is thrown at them. Rather than them just flipping back up onto their feet and getting thrown around again, I want to see them dodging, flipping, and blocking before they get hit. That would blow my mind and could lead to a world of better immersion in games.

[AICoffeeBreak]

Does anyone know how soon this will be used in games? 😅
This is really amazing! Great work from NVIDIA, can't wait to see this technology in the next games I play! Thanks Two Minute Papers for covering this! 🙌

[octopus]

5:55 the cursed movement on the left 😂

[elliotmarks]

I love how, with this video especially, you add humor to these paper demonstrations! It truly makes the field of AI research seem more fun and approachable.