Lava Pit

To cross each bridge, one agent must hold a support platform while the other crosses. Then the roles swap before the team can move on. Progress is monotone and joint.

import comrad

env = comrad.parallel_env("lavapit")

Gymnasium id	comrad/lavapit-v0
Agents	2 by default (supports 2)
Capability tested	Reciprocal assistance, temporary role asymmetry
Episode limit	1000 steps
Observation	RGB image (72, 128, 3) per agent; measurement vector available with additional_input=True
Action space	Tuple(Discrete(3), Discrete(2), Discrete(3))
Default shared_reward	1.0: rewards are fully shared across the team
Score (info["true_objective"])	Joint bridges cleared

Actions

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

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

Reward

One-time frontier milestones, +0.75 per joint platform advance and +2 on success; penalties for stalling, failure, timeout and unfinished bridges.

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

Variants

• None.