MeCo: Enhancing LLM-Empowered Multi-Robot Collaboration via Similar Task Memoization

Codebase for paper: MeCo: Enhancing LLM-Empowered Multi-Robot Collaboration via Similar Task Memoization

Baiqing Wang, Helei Cui, Bo Zhang, Xiaolong Zheng, Bin Guo, Zhiwen Yu

AAMAS 2026 | Arxiv

Setup

setup conda env and package install

conda create -n meco python=3.8
conda activate meco

Install packages

pip install -r requirements.txt

Acquire LLM API Keys

from openai import OpenAI  
client = OpenAI(api_key="", base_url="https://api.deepseek.com")

Usage

LLM-Empowered Method

$ conda activate MeCo
(MeCo) $ python run_llm.py --task A --comm_mode B

A: Select the task you want to execute, including sort, cabinet, rope, sweep, sandwich, pack.

B: Select the execution method: RoCo, Central, HAMS-2, ReAct.

Similarity-aware Method

$ conda activate MeCo
(MeCo) $ python run_similarity_high.py --task A

A: Select the task you want to execute, including sort, cabinet, rope, sweep, sandwich, pack.

There are two versions of the running script: run_similarity_high.py for high-workspace-overlap tasks; run_similarity_low.py for low-workspace-overlap tasks.

This script requires preparing a historical task library and specifying the task library path as shown below:

self.run_dir = os.path.join(base_dir, "path1", "path2", run_name)

MeCo

$ conda activate MeCo
(MeCo) $ python run_meco_high.py --task A --comm_mode B

A: Select the task you want to execute, including sort, cabinet, rope, sweep, sandwich, pack.

B: Select the execution method: RoCo, Central, HAMS-2, ReAct.

There are two versions of the running script: run_meco_high.py for high-workspace-overlap tasks; run_meco_low.py for low-workspace-overlap tasks.

This script requires preparing a historical task library and specifying the task library path as shown below:

self.run_dir = os.path.join(base_dir, "path1", "path2", run_name)

MeCoBench

$ conda activate MeCo
(MeCo) $ python run_mecobench_high.py --task A --mode C

A: Select the task you want to execute, including sort, cabinet, rope, sweep, sandwich, pack.

C: Select the mode of the benchmark you want to generate: similar, dissimilar, random.

There are two versions of the running script: run_mecobench_high.py for high-workspace-overlap tasks; run_mecobench_low.py for low-workspace-overlap tasks.

This script requires preparing a historical task library and specifying the task library path as shown below:

self.run_dir = os.path.join(base_dir, "path1", "path2", run_name)

This script will generate random seeds that are similar/dissimilar/random to the current historical task library.

Contact

Please direct to Baiqing Wang (Email: wbq@mail.nwpu.edu.cn).

Codebase adapted from:

Mandi, Zhao, Shreeya Jain, and Shuran Song. "Roco: Dialectic multi-robot collaboration with large language models." IEEE International Conference on Robotics and Automation (ICRA). 2024. [RoCo Github]