Papers.Data.Code
Repo

mira

mira-wm

Simulates 2v2 Rocket League from all four players' actions and renders synchronized first-person views frame by frame. Includes training code and the aligned 4-player dataset.

GitHub ↗Jul 5, 2026

  • multimodal
  • world-model
  • rocket-league
  • diffusion
Full pipeline diagram showing encoder, latent world model, decoder, and roll-forward loop
Full pipeline diagram showing encoder, latent world model, decoder, and roll-forward loop

What it is

Code and dataset for a real-time 4-player Rocket League world model that generates synchronized views from all players' actions at 20 FPS on one GPU. Run and train a multiplayer world model that simulates full 2v2 Rocket League matches from four players' actions, with synchronized multi-view video generation and a released 4-player dataset.

Key results

  • Runs a 5B latent diffusion world model at 20 FPS on a single GPU
  • Supports full 2v2 matches with 4 synchronized player views and actions
  • Releases 4-second clips with video, actions, events, and physics state

Shows that real-time multiplayer interactive world models are feasible on a single GPU.

One useful ML email a week.

More from Jul 13 – Jul 19

The whole week →

Picked Jul 17

Picked Jul 16

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.