LG is an extension to LaCAM, introducing local guidance — a mechanism that supplies each agent with spatiotemporal hints during multi-agent pathfinding (MAPF). While global guidance provides coarse, static information about congestion mitigation, local guidance dynamically constructs agent-wise short-horizon plans to better handle local bottlenecks and congestion in real-time. The paper will appear at AAAI-26.
LG integrates into the configuration-based MAPF solver LaCAM, maintaining its scalability and completeness while significantly improving solution quality. Empirical results demonstrate that LG reduces flowtime by up to 38% over baseline LaCAM and outperforms state-of-the-art anytime solvers such as LaCAM3 and LNS2, all while preserving real-time responsiveness for up to 1,000 agents.
LaCAM — The baseline configuration-based MAPF solver. |
LG-LaCAM — LaCAM augmented with local guidance. |
Visualization of 400 agents navigating a multi-room environment.
LG-LaCAM visibly alleviates local congestion and accelerates overall flowtime compared to standard LaCAM.
If you find this work to be useful in your research, please consider citing:
@article{arita2025local,
title={Local Guidance for Configuration-Based Multi-Agent Pathfinding},
author={Arita, Tomoki and Okumura, Keisuke},
journal={arXiv preprint arXiv:2510.19072},
year={2025}
}All you need is CMake (≥v3.16). The code is written in C++(17).
First, clone this repo with submodules.
git clone --recursive {this repo}Then, build the project.
cmake -B build && make -C build -j4build/main -i assets/random-32-32-10-random-1.scen -m assets/random-32-32-10.map -N 400 -v 3 --lg --lg_window 20The result will be saved in build/result.txt.
You can find details of all parameters with:
build/main --helpThis repository is compatible with allegorywrite@mapf-visualizer. For example,
mapf-visualizer assets/random-32-32-10.map build/result.txtpre-commit installctest --test-dir ./build

