google-deepmind/lab

Built on the ioquake3 game engine, it exposes navigation and puzzle-solving tasks via a Python API with Lua-scriptable level configuration, supporting both human play and agent training. Renders through EGL/GLX/OSMesa for headless reinforcement learning workflows, and integrates with DeepMind's dm_env framework for standardized RL environment interaction. Tasks range from navigation mazes to multi-agent combat scenarios, all configurable through Lua scripts for domain-specific research variations.

7,338 stars. No commits in the last 6 months.