Neural Network Simulates Rocket League Without Physics Engine

A groundbreaking project named MIRA, developed by Kyutai Labs, Gen Intuition, and Epic Games, has successfully enabled four players to engage in a 2v2 Rocket League match entirely within a neural network. This 5-billion-parameter model generates all visual frames at 20 frames per second using just one Nvidia B200 GPU, remarkably achieving this without any explicit physics or rendering engines. The system learned the game by observing video footage and controller inputs, demonstrating a novel approach to game simulation. The team has open-sourced the code, a technical report, and a substantial 1,000-hour dataset of matches. A notable limitation is the model's short-term memory, retaining information for only about four seconds. Consequently, during replays, the model often hallucinates goals because it has already forgotten the actual event. This project highlights the potential for AI to create complex interactive environments from raw data.