MIRA: Multiplayer Interactive World Models trained on Rocket League [R]
Researchers trained MIRA, a 5-billion-parameter multiplayer interactive world model, on 10,000 hours of synthetic Rocket League gameplay and released a playable demo.
- MIRA is a 5B-parameter multiplayer interactive world model trained on 10,000 hours of synthetic Rocket League data.
- The model runs four-player simulations at 20 fps on a single B200 GPU and includes a playable demo and open dataset.
- Researchers from General Intuition, Kyutai, and Epic Games collaborated on the project.
- The release aims to advance multi-agent AI research with potential applications in robotics and autonomous systems.
A team from General Intuition, Kyutai, and Epic Games has unveiled MIRA, a 5-billion-parameter multiplayer interactive world model trained on 10,000 hours of synthetic Rocket League gameplay. The model simulates four-player interactions in real time, running at 20 frames per second on a single NVIDIA B200 GPU. To foster further research, the team released a playable online demo, a 1,000-hour dataset of four-player gameplay, and an in-depth technical report detailing the architecture and training process. This marks one of the first demonstrations of large-scale, real-time multi-agent world models in a competitive gaming environment, offering a new benchmark for interactive AI systems.
The release highlights the potential of synthetic data for training complex multi-agent models, as Rocket League’s physics-based gameplay provides a controlled yet dynamic environment. By open-sourcing the dataset and demo, the researchers aim to accelerate progress in multiplayer AI, robotics, and autonomous systems where coordinated decision-making is critical. The technical report also explores how such models could scale to more complex scenarios beyond gaming, including real-world simulations and collaborative robotics.
Source: MIRA: Multiplayer Interactive World Models trained on Rocket League [R]. Read the full piece at the source.
Provides a new benchmark for multi-agent world models with open-source tools and datasets.
Demonstrates scalable AI training on synthetic data, relevant for simulation and automation industries.
Highlights growth in multi-agent AI systems with real-time interactive capabilities.
Offers a practical case study in training large-scale world models for competitive environments.
- World Model
- An AI system that simulates and predicts the dynamics of an environment, enabling agents to plan and act within it.
- B200 GPU
- NVIDIA's next-generation GPU designed for high-performance AI and simulation workloads.
Mantle cell lymphoma artificial intelligence prognostic index using hematoxylin and eosin histology - Nature
Claude's 'J-Space' Reveals AI's Internal Reasoning Before Output - 조선일보
