// the take
A physics-based 3D football/soccer simulator from Google Brain, built as a gym-style RL environment on top of an open-source game engine (originally Bastiaan Konings Schuiling's 'Gameplay Football'). It's for RL researchers who want a continuous, high-dimensional control environment with sparse rewards, pixel or vector observations, and a built-in scripted bot opponent — not for people wanting a modern football/sim game.
Ships 50+ pre-built scenarios (academy drills, 11v11, single-goal tasks) that let you test curriculum learning without writing environment code from scratch. Multi-agent support is real, not bolted on — you can control multiple players per side and train against a scripted bot or a checkpointed PPO agent. Offers both raw pixel observations and a structured vector/SMM representation, so you can benchmark CNN policies against simpler MLPs on the same task. The C++ game engine underneath means simulation speed is reasonable for RL rollout volume compared to pure-Python physics environments.
The whole stack is frozen on TensorFlow 1.15 and `dm-sonnet==1.*` — both long dead, so the example PPO training scripts won't run on any modern Python/TF install without real work. Last meaningful code push was mid-2025 but the actual engine and docs haven't moved past the 2019 paper's assumptions; this is a maintained-in-name-only research artifact, not an actively evolved project. Building from source requires compiling a C++ engine with SDL2/boost and platform-specific quirks (Windows needs extra manual steps per the docs) — the Docker path exists because the native build is genuinely finicky. No PyTorch example agents at all, so if your RL stack isn't TF1, you're writing the training loop yourself against their gym-like API.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →