Composite Motion Learning with Task Control

This is the official implementation for

Composite Motion Learning with Task Control [arXiv] [Youtube] (SIGGRAPH'23, TOG)
A GAN-Like Approach for Physics-Based Imitation Learning and Interactive Character Control [arXiv] [Youtube] (SCA'21, PACMCGIT)

We also refer to the extended implementation

AdaptNet: Policy Adaptation for Physics-Based Character Control [arXiv] [Youtube] [webpage] [code] (SIGGRAPH Asia'23, TOG)

Code Usage

Dependencies

Pytorch 1.12
IsaacGym Pr4

We recommend to install all the requirements through Conda by

$ conda create --name <env> --file requirements.txt -c pytorch -c conda-forge

Download IsaacGym Pr4 from the official site and install it via pip.

Policy Training

$ python main.py <configure_file> --ckpt <checkpoint_dir>

We provide our configure files in config folder for reference. To reproduce the examples shown in the paper, e.g. Juggling+Walk, please run the training by

$ python main.py config/juggling+locomotion_walk.py --ckpt ckpt_juggling+locomotion_walk

The training results (model and log) will be generated in the ckpt_juggling+locomotion_walk folder.

The training can be done on a single GPU. Use --device option to specify the device used for training (default: 0). All our results were obtained using machines equipped with Nvidia V100 or A100 GPU.

Policy Evaluation

$ python main.py <configure_file> --ckpt <checkpoint_dir> --test

We provide pretrained policy models in pretrained folder. To evaluate a pretrained policy, e.g. Juggling+Walk, please run

$ python main.py config/juggling+locomotion_walk.py --ckpt pretrained/juggling+locomotion_walk --test

Motion Data Copyright

We provide our motion data in assets/motions.

The data labeled with lafan1 are extracted from Ubisoft LAFAN1 dataset. The juggling motion is extracted from the demo provided by FreeMoCap Project. We cannot provide the tennis motions shown in the paper due to the commercial license.

Citation

If you use the code or provided motions for your work, please consider citing our papers:

@article{composite,
    author = {Xu, Pei and Shang, Xiumin and Zordan, Victor and Karamouzas, Ioannis},
    title = {Composite Motion Learning with Task Control},
    journal = {ACM Transactions on Graphics},
    publisher = {ACM New York, NY, USA},
    year = {2023},
    volume = {42},
    number = {4},
    doi = {10.1145/3592447},
    keywords = {physics-based control, character animation, motion synthesis, reinforcement learning, multi-objective learning, incremental learning, GAN}
}

@article{iccgan,
    author = {Xu, Pei and Karamouzas, Ioannis},
    title = {A GAN-Like Approach for Physics-Based Imitation Learning and Interactive Character Control},
    journal = {Proceedings of the ACM on Computer Graphics and Interactive Techniques},
    publisher = {ACM New York, NY, USA},
    year = {2021},
    volume = {4},
    number = {3},
    pages = {1--22},
    doi = {10.1145/3480148},
    keywords = {physics-based control, character animation, reinforcement learning, GAN}
}