MRBTP: Efficient Multi-Robot Behavior Tree Planning and Collaboration

AAAI 2025 (Oral)

📰 News

• [2026-06] 🎉 MRBTP-demo is now online! A slim, well-documented tutorial version of MRBTP has been released at https://github.com/DIDS-EI/MRBTP-demo. It provides a minimal, runnable entry point for understanding the multi-robot behavior-tree planning algorithm and quickly adapting it to custom tasks.
• [2025-02] 📄 The arXiv preprint is available: https://arxiv.org/abs/2502.18072.
• [2024-12] 🏆 MRBTP is accepted to AAAI 2025 as an Oral presentation (Top 4.6%).

MRBTP is a decentralized Multi-Robot Behavior-Tree Planning framework. Given a shared symbolic goal and per-agent action models, it back-chains runnable behavior trees for each agent so that, executed in parallel, the team collectively satisfies the goal. The repository ships:

• four planning algorithms (MRBTP / MABTP / MAOBTP / CABTP),
• a MiniGrid-based warehouse simulator for quick demos,
• a VirtualHome bridge for household tasks,
• an LLM client used to generate composite-action sub-trees.

---

🛠️ Installation

1. Environment

conda create --name mabtpg python=3.10
conda activate mabtpg

2. Install MABTPG

cd MRBTP
pip install -e .

3. (Optional) VirtualHome executable

Only needed if you want to run the household-service scenario. Currently only Windows is fully tested.

OS	Download
Linux	linux_exec.zip
macOS	macos_exec.zip
Windows	windows_exec.zip

4. MiniGrid / BabyAI

No extra binary download is required — gymnasium and minigrid are pulled in automatically by pip install -e ..

---

🚀 Quick Start

Minimal demo (MiniGrid)

python test_multi_minigrid_single_demo/main.py

Demo with composite actions (LLM-style sub-trees)

test_multi_minigrid_single_demo/main2.py is a self-contained example that combines:

• a custom MiniGrid-DoorKey-8x8-v0 layout,
• 2 agents,
• a CompositeActionPlanner that pre-builds reusable sub-trees (e.g. Move ball-1 to room-1 and place it next to ball-3),
• the optimal planner MAOBTP,
• live rendering of every tick.

python test_multi_minigrid_single_demo/main2.py

All artefacts (robot-i.bt, robot-i.svg, sub-tree diagrams) are written to test_multi_minigrid_single_demo/output/ so the workspace root stays clean.

Customising the environment

Subclass MiniGridToMAGridEnv (see mabtpg/envs/gridenv/minigrid/minigrid_env.py) or register a custom MiniGrid env via gymnasium.envs.registration.register (see the register(...) block in main2.py for a working example) to construct your own room layouts. Pre-built scenarios are listed in MiniGrid_all_scenarios.txt.

---

🧠 Planning Algorithms

The mabtpg/btp/ package exports four classes whose names match the paper one-to-one. See mabtpg/btp/README.md for the full cheat-sheet.

Paper term	Class	File	What it does
MRBTP	MRBTP	multi_robot.py	Top-level facade that returns runnable per-agent BTs.
MR-BTP (baseline)	MABTP	multi_robot_basic.py	Per-step FIFO back-chaining search.
Optimal MR-BTP	MAOBTP	multi_robot_optimal.py	Cost-priority heap search; supports composite actions.
Composite-action BTP	CABTP	composite_action.py	Single-agent planner that builds sub-tree macros.

# Recommended imports
from mabtpg.btp import MRBTP, MABTP, MAOBTP, CABTP
from mabtpg.btp import PlanningAgent, PlanningCondition

# Backward-compatible alias for the paper's historical class name
from mabtpg.btp import DMR     # DMR is MRBTP → True

---

📂 Directory Structure

mabtpg/
│
├── agent/             — Configuration for intelligent agents.
├── llm_client/        — Standalone LLM integration module
│   ├── base.py            BaseLLMClient (chat / tool-calling / embeddings)
│   ├── llms/              Concrete clients (LLMGPT3, LLMGPT4, LLMGPT4o)
│   └── schemas/           Pydantic schemas used as OpenAI tools
├── btp/               — Behavior-tree planning algorithms (see btp/README.md)
│   ├── base/                      PlanningAgent / PlanningCondition
│   ├── multi_robot_basic.py       MABTP   — FIFO baseline
│   ├── multi_robot_optimal.py     MAOBTP  — cost-priority heap; composites
│   ├── composite_action.py        CABTP   — single-agent macro builder
│   └── multi_robot.py             MRBTP   — paper-aligned facade
│                                            (exports DMR = MRBTP alias)
├── behavior_tree/     — Components of the behavior-tree engine
├── envs/              — Environments for agent interaction
│   ├── base/              Foundational elements
│   ├── gridenv/
│   │   └── minigrid/      Warehouse-management scenario
│   ├── virtualhome/
│   │   └── simulation/    VirtualHome Unity / evolving-graph simulators
│   └── numericenv/        Numerical simulation platform
└── utils/             — Supporting functions and utilities

simulators/            — Platforms for realistic training environments

test_multi_minigrid_single_demo/
                       — Smallest end-to-end runnable demo (start here)

test_experiment/       — Modules for testing BT planning, LLMs and scene
│                        interactions.
├── exp1_robustness_parallelism/
│   ├── code/
│   └── results/
└── exp2_subtree_llms/
    ├── code/
    │   ├── data/
    │   └── llm_data/
    └── results/

---

🧪 Reproducing the Paper Experiments

Folder	Topic
test_experiment/exp1_robustness_parallelism/code/	Robustness + parallelism scaling experiments
test_experiment/exp2_subtree_llms/code/	LLM-generated sub-tree macros for composite acts

Each experiment folder ships its own code/ (entry-point scripts) and results/ (CSV / SVG / BT artefacts) so runs are reproducible end-to- end.

---

📑 Citation

If you find this repository useful, please cite:

@inproceedings{mrbtp2025,
  title     = {MRBTP: Efficient Multi-Robot Behavior Tree Planning and Collaboration},
  author    = {Anonymous Authors},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year      = {2025},
  note      = {Oral presentation. arXiv:2502.18072}
}

---

License

This project is licensed under the MIT License — see the LICENSE file for details.