MIRA: a multiplayer world model trained entirely on Rocket League gameplay

MasterScrat built MIRA, a world model that learns the physics and dynamics of Rocket League by training on gameplay data — no hand-crafted rules. What makes it notable for indie AI researchers is the multiplayer angle: most world models tackle single-agent environments, so modeling simultaneous players interacting in a shared space is a harder, less-explored problem.