Analysis updated 2026-05-18
Try the live demo of an AI-generated, playable Rocket League match.
Study how a diffusion based world model can simulate a real time multiplayer game.
Train a single-player or four-player world model on the released Rocket League dataset.
Use the released dataset of synchronized multiplayer gameplay for other research.
| mira-wm/mira | lzh-phd/topic-feasibility-screener | buckyroberts/ella | |
|---|---|---|---|
| Stars | 446 | 446 | 449 |
| Language | Python | Python | Python |
| Last pushed | — | — | 2023-11-09 |
| Maintenance | — | — | Dormant |
| Setup difficulty | hard | moderate | easy |
| Complexity | 5/5 | 3/5 | 2/5 |
| Audience | researcher | researcher | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires an NVIDIA GPU, the pixi package manager, and separately downloading Meta's gated DINOv3 weights for codec training.
MIRA is a research project that builds an AI model able to simulate a video game in real time. Specifically, it recreates Rocket League, a soccer game played with rocket powered cars, using a large diffusion model that generates each video frame on the fly based on what all four players in a match are doing with their controls. A full two versus two match can be played entirely inside this simulated model at twenty frames per second on a single graphics card, and there is a live demo online where people can try it. The project is a joint release from three organizations working together on AI research, and it comes with a technical paper describing how the system was built. Under the hood it combines a video encoder, called a representation autoencoder, with a diffusion model that predicts what the next frame should look like given the players' actions. The repository includes code to install the project using a tool called pixi, which sets up the environment and installs dependencies automatically, though it requires an NVIDIA graphics card. It also provides access to the training dataset used for the project, hosted on Hugging Face, which consists of four second clips from real matches, each with synchronized video from every player, their keyboard inputs, and the underlying game state such as ball position, car positions, and score. There is a small tool to explore this dataset interactively. Beyond just using the released model, the code lets researchers train their own version from scratch, including a single player version and a four player version that builds on it. Training the video encoding step requires a separate frozen model, DINOv3, whose weights must be downloaded from Meta directly under Meta's own terms. This is a research codebase aimed at machine learning researchers and engineers who are comfortable with Python and GPU training pipelines, not a ready-to-use consumer application. The code itself is released under the Apache License 2.0, a permissive open source license that allows free use, including commercial use.
A research AI system that simulates a real time, playable Rocket League match by generating video frames from all four players' controller inputs.
Mainly Python. The stack also includes Python, PyTorch, Hydra.
Released under the Apache License 2.0, which allows free use, modification, and commercial use as long as you keep the license and copyright notice.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.