Lava Maze

A navigator crosses a lava field along a hidden safe path. An observer looks down on the maze and can only help by sending colour signals. The roles have masked action sets.

import comrad

env = comrad.parallel_env("lava_maze")

Gymnasium id	comrad/lava_maze-v0
Agents	2 by default (supports 2)
Capability tested	Emergent communication, asymmetric information
Episode limit	5250 steps
Observation	RGB image (72, 128, 3) per agent; measurement vector available with additional_input=True
Action space	Tuple(Discrete(3), Discrete(3), Discrete(2), Discrete(5))
Default shared_reward	1.0: rewards are fully shared across the team
Score (info["true_objective"])	Maze levels cleared

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. attack: none / attack
4. signal: none / colour 1 / 2 / 3 / 4 (weapon select)

The navigator and the observer have role-specific action masks.

Reward

+10 per cleared level and a bonus for every step closer to the exit; penalties for each step, signals, burns and death.

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

Variants

• lava_maze_simple: smaller, easier maze.
• comrad/scenarios/batch_lava_maze/: pool of generated mazes.