LLM Collaboration – Minecraft

This repo provides the Minecraft environments for CoMLRL.

Installation

Install CoMLRL:

pip install comlrl
# Install PyTorch compatible with your device

Or via conda-forge:

conda install -c conda-forge comlrl
# Install PyTorch compatible with your device

Install the Mineflayer dependencies:

cd LLM_Collab_Minecraft
npm install

Environments

• StrBuild: agents build structures from string blueprints.
• HouseBuild: agents construct houses from layered blueprints under resource limits and spider attacks.

Training Scripts

StrBuild:

python3 str_build/train/train_magrpo.py --config str_build/configs/str_build_magrpo_config.yaml
python3 str_build/train/train_iac.py --config str_build/configs/str_build_iac_config.yaml
python3 str_build/train/train_maac.py --config str_build/configs/str_build_maac_config.yaml

HouseBuild:

python3 house_build/train/train_magrpo.py --config house_build/configs/house_build_magrpo_config.yaml
python3 house_build/train/train_iac.py --config house_build/configs/house_build_iac_config.yaml
python3 house_build/train/train_maac.py --config house_build/configs/house_build_maac_config.yaml

Override any configuration value inline with --override:

python3 str_build/train/train_magrpo.py \
  --config str_build/configs/str_build_magrpo_config.yaml \
  --override model.name='Qwen/Qwen2.5-1.5B-Instruct' magrpo.num_turns=1

Multi-Turn External Feedback

Enable multi-turn training by setting magrpo.num_turns / iac.num_turns / maac.num_turns > 1 and choose an external.mode.

StrBuild modes:

• perfect_feedback
• position_feedback
• score_feedback

HouseBuild modes:

• perfect_feedback
• position_feedback
• position_modification
• rect_modification
• resource_schedule
• score_feedback