This code implements the paper Learning a family of motor skills from a single motion clip. This system learns parameterized motor skills from a single motion clip in physics simulation. Our algorithm can automatically spatio-temporally edit the given motion clip to satisfy given the task space. This code is written in C++ and Python, based on Tensorflow2 and DartSim.
Seyoung Lee, Sunmin Lee, Yongwoo Lee and Jehee Lee. 2021. Learning a family of motor skills from a single motion clip. ACM Trans. Graph. 40, 4. (SIGGRAPH 2021)
Project page: http://mrl.snu.ac.kr/research/ProjectParameterizedMotion/ParameterizedMotion.html
Paper: http://mrl.snu.ac.kr/research/ProjectParameterizedMotion/ParameterizedMotion.pdf
Youtube: https://youtu.be/ihHttL-SeL4
We checked code works in Ubuntu 18.04 and 20.04.
All dependencies can be installed at once by the command below.
sudo ./install.sh
This script download sources on c_env
and py_env
and installs on the same directory.
You can build via:
sudo ./run_cmake.sh
cd build
make -jN
You can render the input motion file/trained network by executing render
file in build/render
.
Sample motions, pretrained model and parameterized model are included in this repository.
You should activate the installed virtual environment by source ../py_env/bin/activate
before running.
You can load BVH files in data/motion
./render --bvh=MOTION_FILE.bvh
You can load trained models in network/output
./render --bvh=MOTION_FILE.bvh --ppo=PRETRAINED_NETWORK/network-0
You can load trained models in network/output
. You can set task parameter and see character tracking the reference generated from elite set or the reference generated from the parameterized network. Elite set data are not included in our sample model.
./render --bvh=MOTION_FILE.bvh --ppo=PARAMETERIZED_NETWORK/network-0 --parametric
Training consists of two stages: pretraining
and parameterized learning
.
Pretraining stage makes the agent able to track the given single motion before parameterized learning.
We use modified version of Deepmimic for pretraining.
Algorithm for parameterized learning is written in our paper.
All possible arguments are written in ppo.py
.
python3 ppo.py --ref=MOTION_FILE.bvh --test_name=TEST_NAME --nslave=N
To load network,
python3 ppo.py --ref=MOTION_FILE.bvh --test_name=TEST_NAME --nslave=N --pretrain=network-0
python3 ppo.py --ref=MOTION_FILE.bvh --test_name=TEST_NAME --nslave=N --pretrain=network-0 --parametric
To load parameterization network and elite set,
python3 ppo.py --ref=MOTION_FILE.bvh --test_name=TEST_NAME --nslave=N --pretrain=network-0 --parametric --load-param
To customize the motion clip and task space, please follow steps below.
You can automatically generate new skeleton that matches input BVH hierarchy by calling function below.
SIM::SkeletonBuilder::generateNewSkeleton(MOTION_PATH, SKELETON_PATH)
The default mass of each body is 5 and you can adjust it by directly editing the generated xml file.
After generating skeleton, you need to edit sim/Configurations.h
to load new xml file and match body names.
Make sure that generated xml file is in data/character
.
You can set the range of task space by editing SIM::EliteSet::setConfigurations()
function.
The parameters you need to edit are
mNumParam
: task dimension
mParamMin
: minimum point in task space
mParamMax
: maximum point in task space
mParamUnit
: diagonal matrix A which is used to calculate distance (See SIM::ParamTree::getDistance()
.)
pBVH
: task parameter of input motion
You can design your own task reward by editing SIM::Controller::getTaskReward()
function.
The final results of parameterized motion depend on the design of fitness function. You can design fitness function by editing SIM::Controller::updateFitness()
function. You may set weight of certain body part higher than other bodies to prevent undesirable deformation of the body.