• [2023.03.28] Due to the cancellation of access to Codex by OpenAI, planning based on Codex is no longer supported by this repository. We will update to the latest OpenAI model, ChatGPT, which has better performance, as soon as possible.

Our codebase require Python ≥ 3.9. Please run the following commands to prepare the environments.

conda create -n planner python=3.9 
conda activate planner
python -m pip install numpy torch==2.0.0.dev20230208+cu117 --index-url https://download.pytorch.org/whl/nightly/cu117
python -m pip install -r requirements.txt
python -m pip install git+https://github.com/MineDojo/MineCLIP

It also requires a modified version of MineDojo as the simulator and a goal-conditioned controller.

git clone https://github.com/CraftJarvis/MC-Simulator.git
cd MC-Simulator
pip install -e .

Below are the configures and weights of models.

Our planner depends on Large Language Model like InstructGPT, Codex or ChatGPT. So we need support the OpenAI keys in the file data/openai_keys.txt. An OpenAI key list is also accepted.

To run the code, call

python main.py model.load_ckpt_path=<path/to/ckpt>

After loading, you should see a window where agents are playing Minecraft.

painting	wooden_slab	stone_stairs

Note: Our planner depends on stable OpenAI API connection. If meeting connection error, please retry it.

Our paper is posted on Arxiv. If it helps you, please consider citing us!

@article{wang2023describe,
  title={Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents},
  author={Wang, Zihao and Cai, Shaofei and Liu, Anji and Ma, Xiaojian and Liang, Yitao},
  journal={arXiv preprint arXiv:2302.01560},
  year={2023}
}