This code implements a basic simulation and control for full-body Musculoskeletal system. Skeletal movements are driven by the actuation of the muscles, coordinated by activation levels. Interfacing with python and pytorch, it is available to use Deep Reinforcement Learning(DRL) algorithm such as Proximal Policy Optimization(PPO).
Seunghwan Lee, Kyoungmin Lee, Moonseok Park, and Jehee Lee Scalable Muscle-actuated Human Simulation and Control, ACM Transactions on Graphics (SIGGRAPH 2019), Volume 37, Article 73.
Project Page : http://mrl.snu.ac.kr/research/ProjectScalable/Page.htm
Youtube : https://youtu.be/a3jfyJ9JVeM
Paper : http://mrl.snu.ac.kr/research/ProjectScalable/Paper.pdf
sudo apt-get install libtinyxml-dev libeigen3-dev libxi-dev libxmu-dev freeglut3-dev libassimp-dev libpython3-dev python3-tk python3-numpy virtualenv ipython3 cmake-curses-gui
We strongly recommand that you install boost libraries from the source code (not apt-get, etc...).
Download boost sources with the version over 1.66.(https://www.boost.org/users/history/version166_0.html)
Compile and Install the sources
cd /path/to/boost_1_xx/ ./bootstrap.sh --with-python=python3 sudo ./b2 --with-python --with-filesystem --with-system --with-regex install
If installed successfully, you should have something like
Please refer to http://dartsim.github.io/ (Install version 6.3)
If you are trying to use latest version, rendering codes should be changed according to the version. It is recommended to use the exact 6.3 version.
Manual from DART(http://dartsim.github.io/installdarton_ubuntu.html) 1. install required dependencies
sudo apt-get install build-essential cmake pkg-config git sudo apt-get install libeigen3-dev libassimp-dev libccd-dev libfcl-dev libboost-regex-dev libboost-system-dev sudo apt-get install libopenscenegraph-dev
git clone git://github.com/dartsim/dart.git cd dart git checkout tags/v6.3.0 mkdir build cd build cmake .. make -j4 sudo make install
You should first activate virtualenv.
bash virtualenv /path/to/venv --python=python3 source /path/to/venv/bin/activate- pytorch(https://pytorch.org/)
pip3 install http://download.pytorch.org/whl/cpu/torch-0.4.1-cp35-cp35m-linux_x86_64.whl pip3 install torchvision
pip3 install numpy matplotlib ipython
Our system require a reference motion to imitate. We provide sample references such as walking, running, and etc...
To learn and simulate, we should provide such a meta data. We provide default meta data in /data/metadata.txt. We parse the text and set the environment. Please note that the learning settings and the test settings should be equal.(metadata.txt should not be changed.)
mkdir build cd build cmake .. make -j8
bash cd python source /path/to/virtualenv/ python3 main.py -d ../data/metadata.txt
All the training networks are saved in /nn folder.
bash source /path/to/virtualenv/ ./render/render ../data/metadata.txt
Run Trained data
bash source /path/to/virtualenv/ ./render/render ../data/metadata.txt ../nn/xxx.pt ../nn/xxx_muscle.pt
If you are simulating with the torque-actuated model,
bash source /path/to/virtualenv/ ./render/render ../data/metadata.txt ../nn/xxx.pt
There is a sample model in data/maya folder that I generally use. Currently if you are trying to edit the model, you have to make your own export maya-python code and xml writer so that the simulation code correctly read the musculoskeletal structure. There is also a rig model that is useful to retarget a new motion.