You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This repository contains the code for evaluating GHIL-Glue with both image-generative and video-generative high-level policies, and implements
our subgoal filtering method.
The code for training the image-generative high-level policies can be found in
the original SuSIE repo. We use the same checkpoints for the image-generative high-level
policies that are used in SuSIE.
The code for training the video-generative high level policies can be found in
our fork of the DynamiCrafter repo.
The code for training the low-level policies and the subgoal classifier networks can be found in our fork of the BridgeData V2 repo.
All checkpoints can be downloaded from https://huggingface.co/kyle-hatch-tri/ghil-glue-checkpoints.
Convert the low-level policy checkpoints to hlo format using external/susie/scripts/policy_ckpt_to_hlo.py
Set the values of the environment variables in external/susie/scripts/robot/eval_diffusion.sh to the paths to your downloaded checkpoints.
Run bash external/susie/scripts/robot/eval_diffusion.sh
Cite
This code is based on calvin-sim from Pranav Atreya.
If you use this code and/or GHIL-Glue in your work, please cite the paper with:
@misc{hatch2024ghilgluehierarchicalcontrolfiltered,
title={GHIL-Glue: Hierarchical Control with Filtered Subgoal Images},
author={Kyle B. Hatch and Ashwin Balakrishna and Oier Mees and Suraj Nair and Seohong Park and Blake Wulfe and Masha Itkina and Benjamin Eysenbach and Sergey Levine and Thomas Kollar and Benjamin Burchfiel},
year={2024},
eprint={2410.20018},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2410.20018},
}