MECoBench

MECoBench is a benchmark for systematically evaluating multimodal multi-agent collaboration in embodied environments.

It contains 192 task cases constructed in VirtualHome, covering two collaboration structures:

• Parallel collaboration: 96 task cases in which agents operate in a shared environment and can complete independent subtasks concurrently.
• Sequential collaboration: 96 task cases in which agents operate in disjoint spatial regions and must coordinate through object handovers.

Dataset Files

File	Description	Number of cases
parallel.json	Parallel collaboration tasks	96
sequential.json	Sequential collaboration tasks	96

Download

Download the complete dataset with:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="q-i-n-g/MECoBench",
    repo_type="dataset"
)

Alternatively, using the Hugging Face CLI:

hf download q-i-n-g/MECoBench \
  --repo-type dataset 

Evaluation

The dataset is designed for use with the MECoBench simulator, which is based on VirtualHome v2.3.0 and includes additional multi-agent actions, agent constraints, rendering capabilities, character assets, and stability fixes.

The simulator, benchmark code, and evaluation scripts are available in the MECoBench GitHub repository.