The DECKARD Minecraft agent uses knowledge from large language models to assist the exploration for reinforcement learning agents. This repository includes our implementation of the DECKARD agent for collecting and crafting arbitrary items in Minecraft. For additional details about our approach please see our website and paper, Do Embodied Agents Dream of Electric Sheep?.

We use Minedojo for agent training and evaluation in Minecraft. Before installing python dependencies for Minedojo, you will need openjdk-8-jdk and python>=3.9. This guide contains additional installation details for the Minedojo simulator.

If you didn't clone the VPT submodule yet, run:

git submodule update --init --recursive

Next, install python packages:

pip install -r requirements.txt

We finetune our agent from OpenAI's VPT Minecraft agent. Download their pretrained weights using our script:

bash download_model.sh

Finally, we use MineClip for reward shaping. Download the weights here and place them at weights/mineclip_attn.pth.

The default way to use DECKARD occasionally pauses exploration to train subtasks. Run this method using:

python main.py

Alternatively, you can pretrain policies for subtasks by running:

python subtask.py --task base_task --target_item log

Then, add the trained subtask checkpoint to your yaml config under techtree_specs.tasks:

my_config:
  task_id: creative
  sim: minedojo
  fast_reset: 0
  terminal_specs:
    max_steps: 10000
  techtree_specs:
    guide_path: data/codex_techtree.json
    target_item: wooden_pickaxe
    tasks:
      log: log_checkpoint.zip

and run DECKARD for building the Minecraft technology tree using:

python techtree.py --config my_config

Note that Minecraft requires using xvfb-run to render on a virtual display when using a headless machine.