MIRA: Multiplayer Interactive World Models Trained on Rocket League

21 pointsposted 2 hours ago
by ethanlipson

5 Comments

jorl17

an hour ago

This was a much better experience than I expected. Rather unbelievable!

Side-effect of the data: clearly the model is better than I normally am at playing, as it spontaneously did several things I had not told it to do and wouldn't really know how to do (at least not with a keyboard).

Really remarkable, congrats!

bschwindHN

27 minutes ago

Where is the option to call all of my tm8s trash? That's an essential part of the experience!

avaer

22 minutes ago

If the data and code is all there, why not release the 5B weights?

MasterScrat

38 minutes ago

Hey all, happy to see this here! This was a colab between General Intuition (that I’m part of), Kyutai and Epic Games.

You can read plenty of details in the blog post and tech report but the TLDR is that we trained a multiplayer world model on 10k hours of Rocket League data. We optimized it to be playable at 20fps on a single GPU.

So what you see in the demo is fully generated: there’s no graphics or physics engine. Instead it’s a 5b neural network that takes actions in and gives pixels out.

superkuh

an hour ago

It feels like playing on a very slow computer. Except that sometimes it just randomly decides you pressed the flip button. Really impressive.