Open-source world model learns Rocket League's physics through multiagent play

Ethan Lipson released MIRA, a world model trained on multiplayer Rocket League footage that can predict game states and agent behavior. The approach sidesteps expensive manual labeling by learning directly from recorded gameplay, potentially offering indie devs a blueprint for building physics-aware AI without massive labeled datasets.