Skip to content

Dataset · kyutai

Multimodal Rocket League dataset for world models

kyutai/rocket-science

Synchronized video, keystrokes, and game state for 4 players from 2v2 Rocket League matches.

What's inside

The dataset contains per-frame game state, game events, keystrokes, and time-synchronized video for all four participants in 2v2 Rocket League matches. The volume ranges from 1 million to 10 million records. The data was collected by Kyutai and General Intuition together with Epic Games to train the MIRA multiplayer world model.

What you can do with it

  • Training generative world models that predict video frames from player actions
  • Training multi-agent game bots using behavioral cloning
  • Researching real-time synchronization of video streams and discrete control signals
Tasks
Other
Languages
English
Size
1M–10M rows
Downloads this month
669
Likes
62
Access
on request on Hugging Face

How to load it

from datasets import load_dataset

ds = load_dataset("kyutai/rocket-science")