This codebase is the official implementation of CurrMask.

Install MuJoCo:

• Download MuJoCo binaries here.
• Unzip the downloaded archive into ~/.mujoco/.
• Append the MuJoCo subdirectory bin path into the env variable LD_LIBRARY_PATH.

Install dependencies in conda environment:

conda env create -f environment.yaml
conda activate currmask

You can follow the example scripts in data_collection/scripts and collect offline data as described in our paper. You can also collect your own dataset following the instructions:

conda activate currmask
cd data_collection
bash scripts/sup.sh             #To collect supervised data
bash scripts/unsup.sh           #To collect unsupervised data

We provide example scritps in folder scripts to pretrain or evaluate the model with skill prompting, goal-conditioned planning and offline RL. An example is:

bash scripts/pretrain.sh
bash scripts/eval.sh

If you find our work helpful, please kindly cite as

@article{tang2024currmask,
      title={Learning Versatile Skills with Curriculum Masking}, 
      author={Yao Tang and Zhihui Xie and Zichuan Lin and Deheng Ye and Shuai Li},
      journal={arXiv preprint arXiv:2410.17744},
      year={2024},
      url={https://arxiv.org/abs/2410.17744}, 
}

This code is built on MaskDP. We would like to express our gratitude to the authors for open-sourcing code to the community!