Scenarios

import comrad

for s in comrad.list_scenarios():
    print(f"{s.name:32} {s.num_agents} agents  score: {s.score}")

Every scenario is a first-person, partially observed, multi-agent task. Each agent sees only its own 128×72 RGB view and acts through a small tuple of discrete action heads. Scenarios are grouped by the cooperative capability they stress.

Scenario	Players	Score (true_objective)	Capability tested
Stag Hunt Arena	2–4	Stags killed	Convention formation, equilibrium selection
Rhythm Sync	2	Stages completed	Temporal coordination
Foraging Commons	2–4	Survival time	Commons dilemma, public goods
Co-op Puzzle	2	Completed puzzle pairs	Turn-taking, partner affordances
Platform Chain	2	Joint checkpoint progress	Long-horizon joint movement
Armory Siege	2–8	Shaped episodic return	Combat coordination, resource routing
Co-op Health Gathering	2–8	Survival time	Resource sharing
Lava Pit	2	Joint bridges cleared	Reciprocal assistance
Smart Enemies	2–4	Enemies killed	Spacing, target prioritisation
Dumb Enemies	2–4	Enemies killed	Cooperative pursuit
Ammo Carrier	2	Survival time	Role-specialised logistics
Stealth Labyrinth	2	Enemies destroyed / total	Extreme role asymmetry
Lava Maze	2	Maze levels cleared	Emergent communication

comrad.list_scenarios() also returns a few easier or alternative variants (platform_chain_easy, lava_maze_simple, rhythm_sync_dense, stag_hunt_arena_continuous*). They are useful for debugging and curricula, but they are not part of the main results table.

Map generation

The maps were generated procedurally. The generators live in the separate DoomGen/ folder of the repository and are not part of the comrad-bench package. The package ships ready-made .wad and .cfg files under comrad/scenarios/.