- Install dependencies. An example
environment.yml
for conda users and an examplerequirements.txt
for pip users are provided. - Download the desired
.weights
file from the VPT github, and place them in a directory calleddata/
. For example:mkdir data
wget -P data/ https://openaipublic.blob.core.windows.net/minecraft-rl/models/bc-house-3x.weights
- Run
run_vpt.py
. Add the--headless
flag for running on headless systems.
- Install dependencies (see above).
- Download MineCLIP weights, STEVE-1 prior weights, and STEVE-1 policy weights:
gdown https://drive.google.com/uc?id=1uaZM1ZLBz2dZWcn85rZmjP7LV6Sg5PZW -O data/mineclip_attn.pth
gdown https://drive.google.com/uc?id=1OdX5wiybK8jALVfP5_dEo0CWm9BQbDES -O data/steve1_prior.pt
gdown https://drive.google.com/uc?id=1E3fd_-H1rRZqMkUKHfiMhx-ppLLehQPI -O data/steve1.weights
- Run
run_steve1.py {prompt}
with the desired prompt- E.g. for the logs task, the prompt 'chop down the tree, gather wood, pick up wood, chop it down, break tree' was used in the paper.
- Add the
--headless
flag for running on headless systems. - Add
--guidance {guidance}
to set the classifier-free guidance scale (default 0).