MirroS

MirroS has released AgentGarten, an environment that connects code-defined physics with a real-time neural renderer. The code handles collisions and state updates, then exports depth and normal geometry conditions to a diffusion renderer generating first-person frames above 30fps, closing the action-observation loop. Agents improve through trial, review, handbook writing, and generational handoff. In a hide-and-seek recreation, they learned to build cover by round 4 and use a ramp to climb a wall by round 10; the team says this is far fewer than the roughly 25 million to 100 million self-play episodes OpenAI needed in 2019. The team also ran four rounds in four other worlds—companion dog, narrow bridge meeting, cooperative sheep herding, and quarry loader.