This repo contains the official implementation of our paper:

Residual Force Control for Agile Human Behavior Imitation and Extended Motion Synthesis
Ye Yuan, Kris Kitani
NeurIPS 2020
[website] [paper] [video]

• The CMU mocap data we use is already included in the data/cmu_mocap folder. The pretrained models are in results/motion_im where each subfolder (e.g., 0506) corresponds to a training config file (e.g., 0506.yml) in motion_imitation/cfg.
• We have provided the following configs (label): 0506 (ballet1), 0507 (ballet2), 0513 (ballet3), 8801 (backflip), 9002 (cartwheel), 9005 (jump kick), 9008 (side flip), 9011 (handspring).

• Tested OS: MacOS, Linux
• Python >= 3.6

1. Install the dependencies:

   pip install -r requirements.txt
   

2. Install MuJoCo following the steps here. Note that mujoco-py (MuJoCo's python binding) is already installed in step 1. This step is to install the actual MuJoCo library.
   Note: MuJoCo is now free thanks to DeepMind! 🎉🎉🎉
3. Set the following environment variable to improve multi-threaded sampling performance:

   export OMP_NUM_THREADS=1
   

python motion_imitation/vis_im.py --cfg 0506 --iter 1000

Here 0506 corresponds to the config motion_imitation/cfg/0506.yml which uses the motion_clip of subject 05 and trial 06 from the CMU mocap database. Check out motion_imitation/cfg for other configs.

Keyboard shortcuts for the GUI:

Note: for motion clips used in the paper, the data has already been processed.

To train policies with other motion clips from the CMU mocap database, you will need to first download the .amc mocap file and place it inside data/cmu_mocap/amc, and run the following script:

python motion_imitation/data_process/convert_cmu_mocap.py --amc_id 05_06 --out_id 05_06_test --render

Here we take 05_06.amc as an example and this command will show a GUI that visualizes the motion clip. The GUI has a few keyboard shortcuts that allow you to adjust the clip and save the training pickle file (05_06_test.p) inside data/cmu_mocap/motion:

If you don't need to adjust the clip, you can just run the script without --render, so the motion pickle is saved without GUI.

You can create your own config file using 0506.yml as a template. Be sure to set motion_id to the motion you have processed. To train an RFC policy, just run:

python motion_imitation/motion_im.py --cfg 0506 --num_threads <max_num_CPU_threads_you_have>

This will save models and logs into results/motion_im/0506.

If you find our work useful in your research, please cite our paper RFC:

@inproceedings{yuan2020residual,
    title={Residual Force Control for Agile Human Behavior Imitation and Extended Motion Synthesis},
    author={Yuan, Ye and Kitani, Kris},
    booktitle={Advances in Neural Information Processing Systems},
    year={2020}
}

The software in this repo is freely available for free non-commercial use. Please see the license for further details.