Install the required dependencies:
pip install -r requirement.txtDownload datasets and pretrained checkpoints using the provided script:
bash download_data.shThis script will download:
- HumanML3D text descriptions and mappings to AMASS motions for generative adaptive sampling
- Pretrained model checkpoints
- Required assets and resources
The data will be saved to sample_data/ and checkpoints to output/HumanoidIm/ directories. Please follow PHC for steps to prepare AMASS training/test datasets.
To train a model, use the following command:
python modskill/run_hydra.py \
learning=<learning_config> \
exp_name=<experiment_name> \
env=<env_config> \
robot=<robot_config> \
env.motion_file=<path_to_motion_file> \
env.num_envs=<num_envs> \
epoch=-1For example:
python modskill/run_hydra.py \
learning=im_pnn_bigger \
exp_name=modskill_attn \
env=env_im_pnn \
robot=smpl_humanoid \
env.motion_file=./sample_data/train.pkl \
env.num_envs=3072 \
epoch=-1To evaluate a trained model, set test=True im_eval=True:
python modskill/run_hydra.py \
learning=<learning_config> \
exp_name=<experiment_name> \
epoch=<checkpoint_epoch> \
test=True \
env=<env_config> \
env.motion_file=<path_to_motion_file> \
env.num_envs=<num_envs> \
env.obs_v=6 \
headless=True \
im_eval=TrueFor example:
python modskill/run_hydra.py \
learning=im_pnn_bigger \
exp_name=modskill_attn \
epoch=-1 \
test=True \
env=env_im_pnn \
env.motion_file=./sample_data/test.pkl \
robot.freeze_hand=True \
robot.box_body=False \
env.num_envs=1 \
env.obs_v=6 \
headless=True \
im_eval=TrueIf you find our work helpful or use our code, please consider citing:
@inproceedings{huang2025modskill,
title={Modskill: Physical character skill modularization},
author={Huang, Yiming and Dou, Zhiyang and Liu, Lingjie},
booktitle={IEEE/CVF International Conference on Computer Vision},
year={2025}
}We would like to acknowledge the following amazing works that this paper builds upon: