// the take
SlimeVolleyGym is a lightweight Gym environment simulating a simple volleyball game, built for testing single-agent, multi-agent, and self-play RL algorithms. It's aimed at RL researchers and students who want a fast, cheap benchmark beyond CartPole/LunarLander but don't need full physics simulation.
Only depends on gym and numpy, so it's unlikely to break from a messy dependency tree. Runs around 12.5K timesteps/sec on a 2015 MacBook for state-space mode, which makes fast iteration possible on a single CPU. The multi-agent API exposes the opponent's observation through info['otherObs'] and normalizes it so a trained agent can swap sides without retraining. It also ships both state-space and pixel observation modes, with the pixel mode tested against standard Atari wrappers, plus pretrained PPO/CMA-ES/GA agents and a tutorial for reproducing results.
Last push was December 2023 and the code targets Gym 0.19-0.20; the author explicitly warns later Gym versions break the API, and there's no Gymnasium port, so plugging this into a current RL stack takes work. Pretrained PPO models were trained on stable-baselines v2.10, not stable-baselines3, so loading them with a modern setup isn't a drop-in. The game physics are admittedly 'dodgy' and the opponent is a single fixed 120-parameter network from 2015, so this is a narrow toy benchmark rather than something that generalizes to harder continuous-control tasks. There's no built-in vectorized environment support — the MPI script for PPO is a one-off, not a general parallelization layer, so throughput on bigger experiments is limited.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →