MIRA AI Generates Real-Time Rocket League Video With Player Input
MIRA shows that a world model trained purely on game data can produce interactive real-time video, and the team's hypothesis about transferring game-trained models to real-world tasks points to a potential path for physical AI.
Reporting from 1 source: GIGAZINE.
Researchers from General Intuition, Kyutai, and Epic Games have released MIRA, an AI model trained on 10,000 hours of Rocket League gameplay. It generates 20 frames per second on a single B200 GPU, allowing players to control a car in real time with no physics or rendering engine. The model uses data from the Necto bot, has a 4-second context window, and the team has released the dataset and code.
MIRA runs at 20 frames per second on a single NVIDIA B200 GPU and can generate a four-player match from every player's perspective simultaneously. A free demo lets anyone control a car using WASD, Q, E, space, O, and P keys, with the model generating the video in real time based on input. The model has no physics or rendering engine, yet it synchronizes all four cars, tracks the boost meter so cars cannot boost when empty, and displays the 'GOAL!' text and effects when a goal is scored.
The development team built MIRA using data from Necto, a machine learning-based Rocket League bot, rather than human play. Because Necto's control precision is high, AI-controlled cars often perform super plays. However, MIRA's context window is only about four seconds, so replays are fabricated randomly and do not match the actual goal. The team describes the project as a stepping stone toward physical AI, hypothesizing that pre-training on large game data and fine-tuning on smaller real-world data could outperform learning from real-world data alone. They have released the full dataset and codebase for training and inference.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.