Stag Hunt Arena

A stag-hunt coordination dilemma. Rabbits give a small, reliable reward to a single hunter. The shielded stag pays much more, but it can only be brought down when the hunters commit to it together. Tests whether policies find the payoff-dominant cooperative equilibrium instead of the safe one.

import comrad

env = comrad.parallel_env("stag_hunt_arena")

Gymnasium id	comrad/stag_hunt_arena-v0
Agents	2 by default (supports 2–4)
Capability tested	Convention formation, trust, equilibrium selection
Episode limit	4500 steps
Observation	RGB image (72, 128, 3) per agent; measurement vector not available
Action space	Tuple(Discrete(3), Discrete(3), Discrete(3), Discrete(2))
Default shared_reward	0.0: each agent receives its own reward
Score (info["true_objective"])	Stags killed

Actions

Each agent sends one choice per head; index 0 is always "do nothing".

1. move: none / forward / backward
2. strafe: none / right / left
3. turn: none / left / right
4. attack: none / attack

Reward

Per agent: +3 per rabbit, +50 per stag, a small bonus for new stag damage (only new lowest stag HP within one stag life); penalties for health loss and death.

The score in info["true_objective"] is what the benchmark reports. The shaped reward is only a training signal.

Variants

• stag_hunt_arena_continuous: continuous turning head.
• stag_hunt_arena_continuous_full: continuous move, strafe and turn heads.