OpenAI's archived multi-agent sandbox

What it does

This is OpenAI’s 2017 multi-agent particle environment: a lightweight 2D physics simulator where multiple agents move, collide, observe, and communicate to solve tasks. It packages nine scenarios—ranging from cooperative navigation to adversarial tag and covert communication—into a Gym-like interface for training multi-agent policies. The codebase is explicitly archived and provided as-is.

The interesting bit

The scenarios are where the personality hides. In “physical deception,” good agents must cover multiple landmarks to fool an adversary about which is the real target; in “covert communication,” two agents try to encrypt a message while a third eavesdrops. It is essentially a minimal physics engine wrapped around a social-dynamics laboratory.

Key highlights

• Nine built-in scenarios mixing cooperation, competition, and communication (including predator-prey, keep-away, and speaker-listener reference games)
• Continuous observation and discrete action spaces with basic 2D interaction physics
• Scenario authoring via four functions: make_world(), reset_world(), reward(), and observation()
• Gym-like API via make_env.py for dropping into RL training loops
• Environment used in the Lowe et al., NIPS 2017 multi-agent actor-critic paper

Caveats

• Explicitly archived with no updates expected; OpenAI points to PettingZoo as the maintained successor
• Dependency versions frozen in 2017 (Python 3.5.4, OpenAI Gym 0.10.5, NumPy 1.14.5), so modern environments will likely need massaging or containerization

Verdict

Grab this if you are reproducing 2017-era multi-agent RL papers or need a minimal, hackable scenario template. Skip it if you want modern Python, current Gymnasium APIs, or active maintenance—use PettingZoo instead.