POGEMA

Partially-Observable Grid Environment for Multiple Agents

POGEMA is a fast, flexible benchmarking platform for cooperative multi-agent pathfinding (MAPF). Agents navigate grid maps under partial observability, making decentralized decisions at each time step.

Installation

pip install git+https://github.com/Cognitive-AI-Systems/pogema.git

Python 3.10+ required. See Installation for extras and development setup.

Highlights

• Fast: Lightweight NumPy-based grid engine — thousands of steps per second
• Flexible: Random or custom maps, 3 collision modes, 3 task modes, configurable observations
• Multi-framework: Native integrations with Gymnasium, PettingZoo, and SampleFactory
• Visualization: Built-in SVG animation recorder with zero overhead when inactive
• Reproducible: Seed-controlled generation for maps, agent positions, and targets

Quick Example

from pogema import pogema_v0, GridConfig

env = pogema_v0(GridConfig(num_agents=4, size=8, seed=42))
obs, info = env.reset()

while True:
    obs, reward, terminated, truncated, info = env.step(env.sample_actions())
    if all(terminated) or all(truncated):
        break

print(info[0]['metrics'])
# {'CSR': 0.0, 'ISR': 0.25, 'EpLength': 64.0, ...}

Citation

If you use POGEMA in your research, please cite:

@inproceedings{skrynnik2025pogema,
  title={POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding},
  author={Skrynnik, Alexey and Andreychuk, Anton and Borzilov, Anatolii and Chernyavskiy, Alexander and Yakovlev, Konstantin and Panov, Aleksandr},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}