# mawpf

**Repository Path**: futurelei/mawpf

## Basic Information

- **Project Name**: mawpf
- **Description**: 从网格世界到仓库：为AGV改编轻量级单次多代理路径寻线，依然来自大神奥村圭介。
他的网站：https://kei18.github.io/#publication
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-22
- **Last Updated**: 2026-06-24

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# MAWPF: Multi-Agent Warehouse Pathfinding

The code repository of the paper **"[From Gridworlds to Warehouses: Adapting Lightweight One-shot Multi-Agent Pathfinding for AGVs](https://arxiv.org/abs/2605.15799)."** (IJCAI-26)

*Multi-agent warehouse pathfinding (MAWPF)* is a gridworld counterpart of MAPF tailored to differential-drive AGVs. Unlike classical MAPF, each agent state encodes not only its grid location but also its *heading* and *speed*, under four warehouse-specific constraints:

1. **Rotation constraints** — agents move in straight lines or rotate in place.
2. **Turning costs** — a 90° rotation takes multiple timesteps ($T_{rot}$).
3. **Acceleration / deceleration** — speed changes by at most one per step, up to $V_{max}$.
4. **Follower collisions** — entering a just-vacated cell is prohibited.

This repository unifies **four representative lightweight MAPF solvers** adapted to MAWPF behind a single executable, selectable with the `--algo` flag:

| `--algo` | Algorithm | Description |
|----------|-----------|-------------|
| `pp`    | PP    | Prioritized Planning with random restart |
| `lns2`  | LNS2  | MAPF Large Neighborhood Search |
| `pibt`  | PIBT  | Multi-step PIBT, rolling horizon (one-shot) |
| `lacam` | LaCAM | LaCAM high-level search with PIBT as the configuration generator |

## Demo

![demo](demo/mawpf.gif)

## Building

All you need is [CMake](https://cmake.org/) (≥ v3.16). The code is written in C++(17).

First, clone this repo.

```sh
git clone https://github.com/HirokiNagai-39/mawpf.git
cd mawpf
```

Then build the project.

```sh
cmake -B build && make -C build
```

## Usage

### Example

```sh
build/main -m maps/random-32-32-20.map -N 50 --algo lacam -t 10 -v 3
```

The result will be saved in `build/result.txt`.

### Selecting the algorithm

Pick one of the four solvers with `-a` / `--algo`:

```sh
build/main -m maps/random-32-32-20.map -N 50 --algo pp     # Prioritized Planning
build/main -m maps/random-32-32-20.map -N 50 --algo lns2   # LNS2
build/main -m maps/random-32-32-20.map -N 50 --algo pibt   # multi-step PIBT
build/main -m maps/random-32-32-20.map -N 50 --algo lacam  # PIBT + LaCAM (default)
```

### Main parameters

- `-m`, `--map` : map file (required)
- `-N`, `--num` : number of agents (required)
- `-a`, `--algo` : algorithm — `pp` | `lns2` | `pibt` | `lacam` (default: `lacam`)
- `-i`, `--scen` : scenario file (optional; random starts/goals are generated otherwise)
- `-s`, `--seed` : random seed
- `-t`, `--time_limit_sec` : time limit in seconds
- `--max_speed` : maximum speed $V_{max}$ (default: 2)
- `--rotation_steps` : timesteps for a 90° rotation $T_{rot}$ (default: 2)
- `--path_length` : PIBT / LaCAM lookahead horizon $L$ (default: 6)

> **Note:** On large maps, a larger $L$ (`--path_length`) makes the pre-processing step take longer.

The format of scenarios and maps follows the [MAPF Benchmarks](https://www.movingai.com/benchmarks/mapf/index.html).

### Help

You can find details of all parameters with:

```sh
build/main --help
```

## Visualizer

A dedicated visualizer for MAWPF solutions (which renders headings, speeds, and rotations in addition to grid positions) will be released soon.

## Licence

This software is released under the MIT License, see [LICENSE](LICENSE).

## Special thanks

[Keisuke Okumura](https://kei18.github.io) — the underlying [lacam3](https://kei18.github.io/lacam3/) framework this work builds upon.