Skip to content

Latest commit

 

History

History
185 lines (147 loc) · 6.94 KB

readme.md

File metadata and controls

185 lines (147 loc) · 6.94 KB

mapf-IR

MIT License

A simulator and visualizer of Multi-Agent Path Finding (MAPF), used in a paper "Iterative Refinement for Real-Time Multi-Robot Path Planning" (to appear at IROS-21). It is written in C++(17) with CMake (≥v3.16) build. The repository uses Google Test and the original library for 2D pathfinding as git submodules. The visualizer uses openFrameworks and works only on macOS.

The implementations include: HCA* and WHCA* [1], PIBT [2], CBS [3], ICBS [4], ECBS [5], Revisit Prioritized Planning [6], Push and Swap [7], winPIBT [8], PIBT+, and IR (Iterative Refinement).

platform status (public) status (dev)
macos-10.15 test_macos build_visualizer_macos test_macos build_visualizer_macos
ubuntu-latest test_ubuntu test_ubuntu

You can see the performance of each solver from auto_record repo. The records were created by Github Actions.

Please cite the following paper if you use the code in your published research:

@inproceedings{okumura2021iterative,
  author={Okumura, Keisuke and Tamura, Yasumasa and Défago, Xavier},
  booktitle={2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  title={Iterative Refinement for Real-Time Multi-Robot Path Planning},
  year={2021},
  pages={9690-9697},
  doi={10.1109/IROS51168.2021.9636071}
}

Demo

100 agents in arena

100 agents in arena, planned by PIBT in 67ms 5ms.

1000 agents in brc202d

1000 agents in brc202d, planned by PIBT in 84sec 1348ms. The gif shows a part of an MAPF plan.

Building

git clone --recursive https://github.com/Kei18/mapf-IR.git
cd mapf-IR
mkdir build
cd build
cmake ..
make

for M1 CPU

cmake -DCPU=M1 ..

Usage

PIBT

./app -i ../instances/sample.txt -s PIBT -o result.txt -v

IR (the result will be saved in result.txt)

./app -i ../instances/random-32-32-20_70agents_1.txt -s IR_HYBRID -n 300 -t 100 -v

You can find details and explanations for all parameters with:

./app --help

Please see instances/sample.txt for parameters of instances, e.g., filed, number of agents, time limit, etc.

Output File

This is an example output of ../instances/sample.txt. Note that (x, y) denotes location. (0, 0) is the left-top point. (x, 0) is the location at x-th column and 1st row.

instance= ../instances/sample.txt
agents=100
map_file=arena.map
solver=PIBT
solved=1
soc=3403
makespan=68
comp_time=58
starts=(32,21),(40,4),(20,22),(26,18), [...]
goals=(10,16),(30,21),(11,42),(44,6), [...]
solution=
0:(32,21),(40,4),(20,22),(26,18), [...]
1:(31,21),(40,5),(20,23),(27,18), [...]
[...]

Visualizer

News

A new visualizer Kei18@mapf-visualizer is available. I recommend using the new one instead of this repo.

Building

It takes around 10 minutes.

macOS

bash ./visualizer/scripts/build_macos.sh

Note: The script of openFrameworks seems to contain bugs. Check this issue. I fixed this in my script :D

Usage

cd build
../visualize.sh result.txt

You can manipulate it via your keyboard. See printed info.

Performance History

Generated by Github Actions. See also auto_record repo.

sub-optimal solvers

optimal solvers

Experimental Environment

v1.1

Scripts for the experiments are in exp_scripts/. Several solvers are coded in different names. See the following.

paper code
RPP RevisitPP
PIBT+ PIBT_COMPLETE
IR: random IR
IR: single-agent IR_SINGLE_AGENTS
IR: focusing-at-goals IR_FOCUS_GOALS
IR: local-repair-around-goals IR_FIX_AT_GOALS
IR: using-MDD IR_MDD
IR: using-bottleneck-agent IR_BOTTLENECK
IR: composition IR_HYBRID

Notes

  • Maps in maps/ are from MAPF benchmarks. When you add a new map, please place it in the maps/ directory.
  • The font in visualizer/bin/data is from Google Fonts.

Licence

This software is released under the MIT License, see LICENSE.txt.

Author

Keisuke Okumura is a Ph.D. student at the Tokyo Institute of Technology, interested in controlling multiple moving agents.

Reference

  1. Silver, D. (2005). Cooperative pathfinding. Proc. AAAI Conf. on Artificial Intelligence and Interactive Digital Entertainment (AIIDE-05)
  2. Okumura, K., Machida, M., Défago, X., & Tamura, Y. (2019). Priority Inheritance with Backtracking for Iterative Multi-agent Path Finding. Proc. Intel. Joint Conf. on Artificial Intelligence (IJCAI)
  3. Sharon, G., Stern, R., Felner, A., & Sturtevant, N. R. (2015). Conflict-based search for optimal multi-agent pathfinding. Artificial Intelligence
  4. Boyarski, E., Felner, A., Stern, R., Sharon, G., Tolpin, D., Betzalel, O., & Shimony, E. (2015). ICBS: improved conflict-based search algorithm for multi-agent pathfinding. Proc. Intel. Joint Conf. on Artificial Intelligence (IJCAI)
  5. Barer, M., Sharon, G., Stern, R., & Felner, A. (2014). Suboptimal Variants of the Conflict-Based Search Algorithm for the Multi-Agent Pathfinding Problem. Annual Symposium on Combinatorial Search (SoCS)
  6. Čáp, M., Novák, P., Kleiner, A., & Selecký, M. (2015). Prioritized planning algorithms for trajectory coordination of multiple mobile robots. IEEE Trans. on automation science and engineering
  7. Luna, R., & Bekris, K. E. (2011). Push and swap: Fast cooperative path-finding with completeness guarantees. Proc. Intel. Joint Conf. on Artificial Intelligence (IJCAI)
  8. Okumura, K., Tamura, Y. & Défago, X. (2020). winPIBT: Extended Prioritized Algorithm for Iterative Multi-agent Path Finding. IJCAI Workshop on Multi-Agent Path Finidng (WoMAPF)