Multi-agent hide-and-seek

A reimplementation of OpenAI's emergent tool-use setup. Hiders get a prep phase to move boxes and ramps, then seekers are released. Agents train with PPO under centralised training and decentralised execution, curriculum from simple arenas to complex ones, and self-play against a pool of past policies. Three arena layouts, 2v2 to 5v5, real-time visualisation with 0.1× to 10× time control. A second environment runs the same idea on MuJoCo worldgen.