pogema 1.1.6

Partially-Observable Grid Environment for Multiple Agents

Partially observable multi-agent pathfinding (PO-MAPF) is a challenging problem which fundamentally differs from regular MAPF, in which a central controller is assumed to construct a joint plan for all agents before they start execution. PO-MAPF is intrisically decentralized and decision making (e.g. planning) here is interleaved with the execution. At each time step an agent receives a (local) observation of the environment and decides which action to take. The ultimate goal for the agents is to reach their goals while avoiding collisions with each other and the static obstacles.

POGEMA stands for Partially-Observable Grid Environment for Multiple Agents. This is a grid-based environment that was specifically designed to be flexible, tunable and scalable. It can be tailored to a variety of PO-MAPF settings. Currently the (somewhat) standard setting is supported: agents can move between the cardinally-adjacent cells of the grid, each action (move or wait) takes one time step. No information sharing between the agents is happening.

POGEMA can generate random maps and start/goals locations for the agents. It also can take custom maps as the input.

Installation

Just install from PyPI:

pip install pogema

Using Example

from pogema import pogema_v0, Hard8x8
env = pogema_v0(grid_config=Hard8x8())

obs = env.reset()

done = [False, ...]

while not all(done):
    # Use random policy to make actions
    obs, reward, done, info = env.step([env.action_space.sample() for _ in range(len(obs))])

Environments

Baselines

The baseline implementations are available as a separate repository.

Interfaces

Pogema provides integrations with a range of MARL frameworks: PettingZoo, PyMARL and SampleFactory.

PettingZoo

from pogema import pogema_v0, GridConfig

# Create Pogema environment with PettingZoo interface
env = pogema_v0(GridConfig(integration="PettingZoo"))

PyMARL

from pogema import pogema_v0, GridConfig

env = pogema_v0(GridConfig(integration="PyMARL"))

SampleFactory

from pogema import pogema_v0, GridConfig

env = pogema_v0(GridConfig(integration="SampleFactory"))

Classic Gym

Pogema is fully capable for single-agent pathfinding tasks.

import gym
import pogema

# This interface provides experience only for agent with id=0,
# other agents will take random actions.
env = gym.make("Pogema-v0")

Example of training stable-baselines3 DQN to solve single-agent pathfinding tasks:

Customization

Random maps

from pogema import pogema_v0, GridConfig

# Define random configuration
grid_config = GridConfig(num_agents=4,  # number of agents
                         size=8, # size of the grid
                         density=0.4,  # obstacle density
                         seed=1,  # set to None for random 
                                  # obstacles, agents and targets 
                                  # positions at each reset
                         max_episode_steps=128,  # horizon
                         obs_radius=3,  # defines field of view
                         )

env = pogema_v0(grid_config=grid_config)
env.reset()
env.render()

Custom maps

from pogema import pogema_v0, GridConfig

grid = """
.....#.....
.....#.....
...........
.....#.....
.....#.....
#.####.....
.....###.##
.....#.....
.....#.....
...........
.....#.....
"""

# Define new configuration with 8 randomly placed agents
grid_config = GridConfig(map=grid, num_agents=8)

# Create custom Pogema environment
env = pogema_v0(grid_config=grid_config)

Citation

If you use this repository in your research or wish to cite it, please make a reference to our paper:

@misc{https://doi.org/10.48550/arxiv.2206.10944,
  doi = {10.48550/ARXIV.2206.10944},  
  url = {https://arxiv.org/abs/2206.10944},
  author = {Skrynnik, Alexey and Andreychuk, Anton and Yakovlev, Konstantin and Panov, Aleksandr I.},
  keywords = {Machine Learning (cs.LG), Artificial Intelligence (cs.AI), Multiagent Systems (cs.MA), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {POGEMA: Partially Observable Grid Environment for Multiple Agents},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}