Skip to content

COMRAD: Cooperative multi-agent reinforcement learning from pixels.

🎮 Project Page GitHub PyPI

COMRAD is a benchmark of 13 first-person cooperative scenarios built on ViZDoom. Each scenario isolates one coordination skill: committing to a joint plan, synchronising in time, sharing a commons, swapping roles, or communicating without words.

COMRAD agents playing

pip install comrad-bench

import comrad

env = comrad.parallel_env("stag_hunt_arena")   # PettingZoo ParallelEnv
observations, infos = env.reset(seed=42)
while env.agents:
    actions = {agent: env.action_space(agent).sample() for agent in env.agents}
    observations, rewards, terminations, truncations, infos = env.step(actions)

Works with any MARL library that speaks PettingZoo or Gymnasium. No Sample Factory configuration is needed.

Scenarios
Stag Hunt Arena
Stag Hunt Arena		 Rhythm Sync
Rhythm Sync		 Foraging Commons
Foraging Commons
Co-op Puzzle
Co-op Puzzle		 Platform Chain
Platform Chain		 Armory Siege
Armory Siege
Co-op Health Gathering
Co-op Health Gathering		 Lava Pit
Lava Pit		 Smart / Dumb Enemies
Smart / Dumb Enemies
Stealth Labyrinth
Stealth Labyrinth		 Ammo Carrier
Ammo Carrier		 Lava Maze
Lava Maze

Where to next