Research map / Multi-agent systems and games / Cooperative multi-agent RL
POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces POGEMA, a Python toolkit (environment, map generator, visualizer, distributed benchmarking toolbox, and evaluation protocol with derived metrics) for benchmarking learning-based, hybrid, and search-based multi-agent pathfinding methods.
Problem
Fair comparison of classical, learning-based, and hybrid multi-agent pathfinding (MAPF/LMAPF) approaches is hindered by the lack of a unified environment, instance sets, and evaluation protocol.
Approach
A fast, pure-Python grid environment with procedural map and instance generation, a parallel benchmarking toolbox, and a protocol of high-level metrics (performance, out-of-distribution, cooperation, scalability, coordination, pathfinding) computed from SoC, makespan, throughput and success rate, used to compare 12 MARL, hybrid, and planning baselines.
Key results
- In MAPF, centralized LaCAM achieves CSR 1.00 and SoC 20.84 on Puzzles, while the best learnable method SCRIMP reaches CSR 0.85.
- In LMAPF on Puzzles, centralized RHCR achieves average throughput 0.538, versus 0.394 for the best learnable method MATS-LP.
- All MARL baselines (MAMBA, QMIX, QPLEX, VDN, IQL) fail to solve any Cities-tiles out-of-distribution instances in MAPF, while hybrid and search methods perform better.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LaCAM instead of other evaluated approaches | MAPF solvers · Puzzles | CSR: 1.00 ± 0.00 | Helped |
| Use SCRIMP instead of DCC | learnable MAPF solvers · Puzzles | CSR: 0.74 ± 0.04; 0.85 ± 0.03 | Helped |
| Use MAMBA instead of other MARL approaches | MAPF solvers · Random and Mazes | performance | Helped |
| Use MAMBA for out-of-distribution MAPF | MARL solvers · Cities-tiles | success rate | Hurt |
| Use MAMBA rather than the other evaluated approaches | MAPF solvers · Cities (single-agent pathfinding) | makespan: 416.45 ± 139.34 | Hurt |
| Use MATS-LP instead of Follower | LMAPF solvers · Puzzles | average throughput: 0.319 ± 0.020; 0.394 ± 0.021 | Helped |
| Use MATS-LP rather than Follower | LMAPF solvers · Cities-tiles, 64, 128 and 192 agents | throughput | Helped |
| Use Follower rather than MATS-LP | LMAPF solvers · Cities-tiles, 256 agents | throughput | Helped |
| Use RHCR instead of decentralized approaches | LMAPF solvers · Puzzles | average throughput: 0.538 ± 0.021 | Helped |
| Use MAMBA rather than other MARL approaches | LMAPF solvers · Random and Mazes | throughput | Hurt |
| Use the Switcher approaches rather than MARL approaches | LMAPF solvers · Random and Mazes | collisions | Helped |
| Use LaCAM rather than SCRIMP on single-agent pathfinding | MAPF solvers · Cities | makespan: 179.82 ± 20.97; 755.98 ± 183.31 | Helped |
Surprising findings
- SCRIMP underperforms on single-agent pathfinding despite strong multi-agent results, suggesting a weakness in scenarios without communication need.
- ASwitcher outperforms several MARL methods (MAMBA, QMIX, IQL, VDN) that are given a global map on pathfinding.
Limitations
- Communication between agents is not currently disentangled in the environment.
- No JAX support and no integration with GPU parallelization tools.
- MARL baselines perform far below specialized hybrid and search-based methods, possibly due to missing techniques.
Experimental setup
Datasets: MovingAI city maps procedurally generated random, maze, and warehouse maps hand-crafted Puzzles maps
Benchmarks: POGEMA benchmark (Mazes, Random, Warehouse, Puzzles, Cities, Cities-tiles)
Metrics: success rate (CSR) sum-of-costs makespan throughput Performance Out-of-Distribution Cooperation Scalability Coordination Pathfinding
Baselines: MAMBA QPLEX IQL VDN QMIX LaCAM RHCR SCRIMP DCC Follower MATS-LP Switcher
Builds on: PettingZoo PyMARL2 Gymnasium PPO Follower MATS-LP
Closest papers by meaning
- LNS2+RL: Combining Multi-agent Reinforcement Learning with Large Neighborhood Search in Multi-agent Path Finding (AAAI 2025 · similarity 0.89)
- Learn to Follow: Decentralized Lifelong Multi-Agent Pathfinding via Planning and Learning (AAAI 2024 · similarity 0.89)
- MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale (AAAI 2025 · similarity 0.88)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.88)
- Multiagent Gumbel MuZero: Efficient Planning in Combinatorial Action Spaces (AAAI 2024 · similarity 0.87)
- Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.