Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale
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.
Trains transformer policies (2M-85M params) by imitation learning on 1B observation-action pairs from LaCAM expert MAPF solutions, yielding a decentralized, communication-free MAPF solver (MAPF-GPT) that outperforms DCC and SCRIMP.
Problem
Learning-based multi-agent pathfinding usually relies on RL plus extra modules (single-agent planning, communication); the question is whether purely supervised imitation learning at scale on expert MAPF solutions can yield a strong decentralized solver that generalizes out of distribution.
Approach
Tokenize each agent's local observation (cost-to-go field of view, nearby agents' positions, goals, action history, greedy actions) into a 67-token vocabulary and train a non-autoregressive decoder-only transformer with cross-entropy to imitate LaCAM's actions, producing a foundation-style MAPF policy with no heuristics or communication.
Key results
- MAPF-GPT variants outperform DCC and SCRIMP on Random and Mazes maps in success rate; on Warehouse SCRIMP is superior with more than 128 agents.
- Ablation on MAPF-GPT-6M: removing cost-to-go information drops Random success from 97.6% to 25.8%, Mazes from 74.6% to 15.1%.
- MAPF-GPT-2M and 6M are more than 13x faster than SCRIMP and 8x faster than DCC at 192 agents on Warehouse.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune MAPF-GPT-6M on LifeLong MAPF expert data | MAPF-GPT-6M · LifeLong MAPF, Random | throughput: 1.497 -> 1.507 | Helped |
| fine-tune MAPF-GPT-6M on LifeLong MAPF expert data | MAPF-GPT-6M · LifeLong MAPF, Mazes | throughput: 0.908 -> 1.087 | Helped |
| fine-tune MAPF-GPT-6M on LifeLong MAPF expert data | MAPF-GPT-6M · LifeLong MAPF, Warehouse | throughput: 1.113 -> 1.270 | Helped |
| fine-tune MAPF-GPT-6M on LifeLong MAPF expert data | MAPF-GPT-6M · LifeLong MAPF, Cities-tiles | throughput: 2.840 -> 2.994 | Helped |
| add goal information for all agents (ablation) | MAPF-GPT-6M · Random, Mazes, Warehouse, and Cities-tiles | success rate: 95.7% -> 97.6%; 71.6% -> 74.6%; 92.8% -> 94.1%; 88.4% -> 82.0% | Mixed |
| add greedy-action information (ablation) | MAPF-GPT-6M · Random, Mazes, Warehouse, Cities-tiles, and Puzzles | success rate: 97.0% -> 97.6%; 37.6% -> 74.6%; 87.7% -> 94.1%; 79.1% -> 82.0%; 92.7% -> 94.0% | Helped |
| add action history (ablation) | MAPF-GPT-6M · Random, Mazes, Warehouse, Cities-tiles, and Puzzles | success rate: 95.6% -> 97.6%; 85.8% -> 74.6%; 94.8% -> 94.1%; 82.2% -> 82.0%; 91.5% -> 94.0% | Mixed |
| add cost-to-go information (ablation) | MAPF-GPT-6M · Random, Mazes, Warehouse, Cities-tiles, and Puzzles | success rate: 25.8% -> 97.6%; 15.1% -> 74.6%; 11.5% -> 94.1%; 10.2% -> 82.0%; 52.5% -> 94.0% | Helped |
Surprising findings
- Removing action history improved performance on Mazes (85.8% vs 74.6%) and Warehouse (94.8% vs 94.1%), suggesting action history is not crucial for behavioral cloning.
- Removing goal information improved Cities-tiles success (88.4% vs 82.0%).
Limitations
- No theoretical guarantees, shared with learnable methods.
- Training the 85M model is computationally demanding (243 hours on 4x H100).
- Performance is sensitive to expert data quality, and imitation of non-LaCAM centralized experts such as CBS is unexplored.
Experimental setup
Datasets: MAPF expert dataset generated with LaCAM on POGEMA maze and random maps
Benchmarks: POGEMA Random POGEMA Mazes POGEMA Warehouse POGEMA Cities-tiles POGEMA Puzzles Lifelong MAPF throughput
Metrics: success rate sum-of-costs relative to LaCAM throughput runtime per action
Baselines: DCC SCRIMP LaCAM RHCR Follower MATS-LP
Builds on: LaCAM NanoGPT POGEMA Decision Transformer
Closest papers by meaning
- Learn to Follow: Decentralized Lifelong Multi-Agent Pathfinding via Planning and Learning (AAAI 2024 · similarity 0.91)
- Decentralized Monte Carlo Tree Search for Partially Observable Multi-Agent Pathfinding (AAAI 2024 · similarity 0.90)
- LNS2+RL: Combining Multi-agent Reinforcement Learning with Large Neighborhood Search in Multi-agent Path Finding (AAAI 2025 · similarity 0.89)
- Traffic Flow Optimisation for Lifelong Multi-Agent Path Finding (AAAI 2024 · similarity 0.89)
- POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding (ICLR 2025 · similarity 0.88)
- Anytime Multi-Agent Path Finding with an Adaptive Delay-Based Heuristic (AAAI 2025 · 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.