Melting Pot — Multi-Agent RL Test Suite

Melting Pot — Multi-Agent RL Test Suite

Melting Pot is a Google DeepMind research suite with 50+ multi-agent social scenarios designed to evaluate generalization in MARL — can an agent trained in one social context transfer to novel partners and situations? Scenarios cover cooperation dilemmas, competitive games, resource harvesting, and commons management.

Key features

• 50+ scenarios across cooperation, competition, mixed motives, and signaling
• Population-based evaluation protocol for measuring generalization (not just peak performance)
• Built on DeepMind Lab2D for grid-world environments
• Reference baseline agents (MEME, A3C) and evaluation metrics included
• Python API compatible with RLlib, ACME, and custom training loops

Quick start

pip install dm-meltingpot

import meltingpot

env = meltingpot.substrate.build("commons_harvest__open",
                                  roles=["default"] * 5)
step = env.reset()
for _ in range(500):
    actions = [env.action_space[i].sample() for i in range(5)]
    step = env.step(actions)
env.close()

npx ai-supply add meltingpot-multiagent-rl

Curated mirror of the open-source Melting Pot (Apache-2.0). Get it from the source.