conda create -n humanoid python=3.8
conda activate humanoid
pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
git clone git@github.com:jimazeyu/exbody2.git
cd exbody2
# Download the Isaac Gym binaries from https://developer.nvidia.com/isaac-gym
cd isaacgym/python && pip install -e .
cd ~/exbody2/rsl_rl && pip install -e .
cd ~/exbody2/legged_gym && pip install -e .
pip install "numpy<1.24" pydelatin wandb tqdm opencv-python ipdb pyfqmr flask dill gdownI put the motions_amass_CMU.yaml file in ASE/ase/poselib/data/configs as an example. The .npy files can be gotten from method2. If you want to use other training motions, you should edit your own yaml configs. You may skip the data preparation and test the code with the example motions in the motions_dance_release.pkl.
- Install FBX SDK (see: nv-tlabs/ASE#61).
- Follow Expressive Humanoid to retarget .fbx files to .npy motion format.
- Use amass_g1_retargeting to convert AMASS data into .npy motions.
- Place the output .npy files into ASE/ase/poselib/data/g1_retarget_npy/.
- Follow Expressive Humanoid to generate key points.
python train.py --task g1_mimic_priv 000-00-some_description --motion_name motions_dance_release.yaml --device cuda:0 --entity WANDB_ENTITYpython train.py --task g1_mimic_priv_distill 000-01-student_description --motion_name motions_dance_release.yaml --device cuda:0 --entity WANDB_ENTITY --resume --resumeid 000-00python play_priv.py --task g1_mimic_priv 000-00 --motion_name motions_dance_release.yaml --device cuda:0python play_priv.py --task g1_mimic_priv_distill 000-01 --motion_name motions_dance_release.yaml --device cuda:0motions_dance_release.yaml corresponds to motions_dance_release.pkl, which contains 8 example dance motions. These are preprocessed and included in the repo for direct training and testing.
Warning: the example dancing motions are not easy, you may need really good g1 to try. I suggest you start with simple motions!!!
For more viewer operations, debugging options, and experiment configurations, please refer to the Expressive Humanoid instructions — most Isaac Gym controls are consistent.
- Release the example policy.
- Release the complete training pipeline.
- Release the deployment code.
- Retargeting and motion processing code is adapted from ASE and PHC.
- The codebase adapted from Expressive Humanoid.