A new AI model named MIRA, or Multiplayer Interactive World Models with Representation Autoencoders, has demonstrated a significant leap in the field of artificial intelligence by learning a consistent representation of the physics, 3D environment, and game state of the popular video game Rocket League. The project, a collaborative effort between General Intuition, Kyutai, and Epic Games, was highlighted by Michael Black, an author on the project, who expressed excitement about its potential for real-world generalization.
The MIRA model is a 5-billion-parameter diffusion transformer paired with a 600-million-parameter video representation codec. It can simulate 2v2 Rocket League matches in real-time at 20 frames per second, responding to player actions and generating a consistent view across four players. Notably, the model operates without an explicit physics engine or 3D representation, learning solely from video and action data.
"The term 'world model' is tossed around a lot but this is the real deal," Michael Black stated in the tweet. He added, "I'm excited to have been part of this project because I see this as a step towards learning world models that can generalized to the real world."
The model was trained on approximately 10,000 hours of 2v2 Rocket League matches generated entirely by bots, not human players. This approach leverages the abundance and cleanliness of synthetic data, a methodology Michael Black has long championed through projects like Sintel, SURREAL, AGORA, BEDLAM, and BEDLAM2. The use of bot-generated data provides a scalable pathway for training robust world models.
MIRA's development is seen as a crucial stepping stone towards physical AI applications such as robotics and self-driving cars. By first mastering complex virtual environments, researchers aim to develop models that can transfer learned behaviors to the messier and more data-scarce real world. The project's dataset, code, and technical report have been open-sourced, encouraging further research and development in the field.
