IngyN / gym-grid-ma

Gym for 2d maze with configurable targets and multiple agents

Grid World Gym

simple multi-agent gridworld gym loaded from text file maps with simple collision detection and prevention. This is the environment used for the gridworld environment in Safe multi-agent reinforcement learning via shielding.

Relevant repositories:

• Shielded MARL
• Shielded Deep MARL

Notes:

• Step function:
  • The reward for a collision can be set where collision_cost = -reward. If you want the reward of a collision to be -10 then collision_cost=10.
  • Penalty for unmoving agents can be set by having noop=True
  • The share variable should be True for shared goals (ie. if multiple agents are going to one goal location).
  • Out of bounds checking with penalty -10.
  • Random priority of agents for conflicting agent actions can be enabled with random_priority=True.

Requires the OpenAI gym package.