Cognitive-AI-Systems/pogema
[ICLR-2025] 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 / MAPF settings.
What it solves
POGEMA provides a standardized, flexible environment for testing and developing algorithms for Partially-Observable Multi-Agent Pathfinding (PO-MAPF). Unlike traditional pathfinding where a central controller plans for all agents, PO-MAPF requires agents to make decentralized decisions based on local observations to reach their goals while avoiding collisions with other agents and obstacles.
How it works
It is a grid-based simulation environment where agents move between cardinal-adjacent cells. Each agent operates independently without information sharing, relying on a local field of view (observation radius) to navigate. The system can generate random maps with tunable parameters (like obstacle density and grid size) or load custom map layouts.
Who it’s for
Researchers and developers working on multi-agent reinforcement learning (MARL) and decentralized pathfinding algorithms.
Highlights
- MARL Integration: Native support for popular frameworks including PettingZoo, PyMARL, SampleFactory, and Gymnasium.
- Customizable Environments: Ability to generate random grids or use custom text-based map definitions.
- Scalable Design: Tunable parameters for agent count, grid size, and observation radius to test scalability.
- Detailed Evaluation: Accompanied by a dedicated benchmark repository for baseline implementations and evaluation protocols.
Related
- Project
- Project
- Project
- Project