Motion synthesis in a dynamic environment has been a long-standing problem for character animation. Methods using motion capture data tend to scale poorly in complex environments because of their larger capturing and labeling requirement. Physics-based controllers are effective in this regard, albeit less controllable. In this paper, we present CARL, a quadruped agent that can be controlled with high-level directives and react naturally to dynamic environments. Starting with an agent that can imitate individual animation clips, we use Generative Adversarial Networks to adapt high-level controls, such as speed and heading, to action distributions that correspond to the original animations. Further fine-tuning through the deep reinforcement learning enables the agent to recover from unseen external perturbations while producing smooth transitions. It then becomes straightforward to create autonomous agents in dynamic environments by adding navigation modules over the entire process. We evaluate our approach by measuring the agent's ability to follow user control and provide a visual analysis of the generated motion to show its effectiveness.

Project page: https://inventec-ai-center.github.io/projects/CARL/index.html
Youtube: https://youtu.be/t9CdF_Pl19Q
Paper: https://arxiv.org/abs/2005.03288

1. Install dependencies

sudo apt-get update
sudo apt-get install -y vim cmake wget git tar unzip build-essential clang llvm python-dev freeglut3-dev libglew-dev swig python3-pip

2. Clone the repository

git clone --recursive https://github.com/inventec-ai-center/carl-siggraph2020.git

3. Install Eigen (version: 3.3.7)

cd ~
wget https://gitlab.com/libeigen/eigen/-/archive/3.3.7/eigen-3.3.7.tar.bz2
tar -xf eigen-3.3.7.tar.bz2
cd eigen-3.3.7
mkdir build_cmake
cd build_cmake
cmake ..
sudo make install

4. Install Bullet (version: 2.88)

cd ~
wget https://github.com/bulletphysics/bullet3/archive/2.88.zip
unzip 2.88.zip
cd bullet3-2.88/
vim build_cmake_pybullet_double.sh
# replace -DUSE_DOUBLE_PRECISION=ON with -DUSE_DOUBLE_PRECISION=OFF
./build_cmake_pybullet_double.sh
cd build_cmake
sudo make install

5. Install Open MPI (version: 1.10.7)

cd ~
wget https://download.open-mpi.org/release/open-mpi/v1.10/openmpi-1.10.7.tar.gz
gunzip -c openmpi-1.10.7.tar.gz | tar xf -
cd openmpi-1.10.7
./configure --prefix=/usr/local
sudo make all install -j8
ldconfig

6. Install Miniconda

cd ~
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod 777 Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh
# restart bash to activate conda
conda create --name env_carl python=3.7
conda activate env_carl

7. Install python packages

cd {CARL_REPO}/CARL # {CARL_REPO} is the path where the CARL repository is located
pip install --upgrade pip
pip install -r requirements.txt

8. Modify the Makefile in CARL/CarlCore/ by specifying the following

   • EIGEN_DIR: Eigen include directory
   • BULLET_INC_DIR: Bullet source directory
   • PYTHON_INC: python include directory
   • PYTHON_LIB: python lib directory

9. Compile CarlCore

cd {CARL_REPO}/CARL/CarlCore # {CARL_REPO} is the path where the CARL repository is located
make python -j8

For convenience, we include the pretrained models under data/policies/dog3d and reference motions under --motion_files. To run each training stages please follow the instructions below.

• Visualize the Reference Motion (you can modify the reference motion by setting the --motion_files in the config file run_dog3d_reference_motion_args.txt)

python CARL_Runner.py --arg_file args/run_dog3d_reference_motion_args.txt

• Stage 1: Imitation Learning

# for speed control
python CARL_Runner.py --arg_file args/run_dog3d_imitation_learning_speed_control_args.txt

# for heading control
python CARL_Runner.py --arg_file args/run_dog3d_imitation_learning_heading_control_args.txt

• Stage 2: GAN Control Adapter

# for speed control
python CARL_Runner.py --arg_file args/run_dog3d_gan_control_adapter_speed_control_args.txt

# for heading control
python CARL_Runner.py --arg_file args/run_dog3d_gan_control_adapter_heading_control_args.txt

• Stage 3: DRL Fine-tuning

# for speed control
python CARL_Runner.py --arg_file args/run_dog3d_drl_finetuning_speed_control_args.txt

# for heading control
python CARL_Runner.py --arg_file args/run_dog3d_drl_finetuning_heading_control_args.txt

• Stage 1: Imitation Learning

# for speed control
python mpi_run.py --arg_file args/train_dog3d_imitation_learning_speed_control_args.txt --num_workers 8

# for heading control
python mpi_run.py --arg_file args/train_dog3d_imitation_learning_heading_control_args.txt --num_workers 8

• Stage 2: GAN Control Adapter

# for speed control
## Step 1: get training samples drawn from real distribution
python mpi_run.py --arg_file args/generate_dog3d_gan_control_adapter_samples_speed_control_args.txt --num_workers 8
## Step 2: train the GAN control adapter
python CARL_Train_Control_Adapter.py --arg_file args/train_dog3d_gan_control_adapter_speed_control_args.txt

# for heading control
python mpi_run.py --arg_file args/generate_dog3d_gan_control_adapter_samples_heading_control_args.txt --num_workers 8
python CARL_Train_Control_Adapter.py --arg_file args/train_dog3d_gan_control_adapter_heading_control_args.txt

• Stage 3: DRL Fine-tuning

# for speed control
python mpi_run.py --arg_file args/train_dog3d_drl_finetuning_speed_control_args.txt --num_workers 8

# for heading control
python mpi_run.py --arg_file args/train_dog3d_drl_finetuning_heading_control_args.txt --num_workers 8

We provided a jupyter notebook file plot_action_dist.ipynb and a python script file plot_action_dist.py to visualize the action distribution. Visit the file for more information.

# generate real samples drawn from real distribution
python mpi_run.py --arg_file args/generate_dog3d_gan_control_adapter_samples_speed_control_args.txt --num_workers 8

# save action distribution image
python plot_action_dist.py

Note: You'll need to install python package jupyter before using the jupyter notebook file plot_action_dist.ipynb.

• right-click and drag will pan the camera
• scroll wheel will zoom in/out
• pressing 'r' will reset the episode
• pressing 'l' will reload the argument file and rebuild everything
• pressing 'g' will start/stop video recording (required ffmpeg)
• pressing space will pause/resume the simulation
• pressing '>' will step the simulation one step at a time

• pressing 'u' will show/hide gait patterns
• pressing 'k' will show/hide character's shapes

• pressing 'k' will show/hide the reference motion
• pressing 'x' will pelt the character with random boxes
• left-click and drag will apply a force on the character at a particular location

• pressing 'a' will increase the target speed
• pressing 'd' will decrease the target speed
• pressing 'q' will drastically increase the target speed
• pressing 'e' will drastically decrease the target speed
• pressing 's' will switch modes between manual and random controls of the target speed
• pressing 'x' will pelt the character with random boxes
• left-click and drag will apply a force on the character at a particular location

• pressing 'a' will increase the target heading
• pressing 'd' will decrease the target heading
• pressing 'q' will drastically increase the target heading
• pressing 'e' will drastically decrease the target heading
• pressing 's' will switch modes among manual, random and pre-defined path controls of the target heading
• pressing 'x' will pelt the character with random boxes
• left-click and drag will apply a force on the character at a particular location

We use a readily available kinematic controller called Mode Adaptive Neural Network (MANN) by [Zhang et al. 2018] (visit their repository at https://github.com/sebastianstarke/AI4Animation for more detail) to get the reference motions. The system doesn't require such specific animation system to work. Motion graphs or motion capture data are also viable alternatives to use.

For your convenience, we provided reference motions for both speed and heading control tasks located in data/motions/.

Inspired by DeepMimic, the motion files follow the JSON format. The "Loop" field specifies whether or not the motion is cyclic. "wrap" specifies a cyclic motion that will wrap back to the start at the end, while "none" specifies an acyclic motion that will stop once it reaches the end of the motion. Each vector in the "Frames" list specifies a keyframe in the motion. Each frame has the following format:

[
    duration of frame in seconds (1D),
    root position (3D),
    root rotation (4D),
    hips rotation (4D),
    spine rotation (4D),
    spine1 rotation (4D),
    neck rotation (4D),
    head rotation (4D),
    head site rotation (4D),
    left shoulder rotation (4D),
    left arm rotation (4D),
    left fore arm rotation (4D),
    left hand rotation (4D),
    left hand side rotation (4D),
    right shoulder rotation (4D),
    right arm rotation (4D),
    right fore arm rotation (4D),
    right hand rotation (4D),
    right hand side rotation (4D),
    left up leg rotation (4D),
    left leg rotation (4D),
    left foot rotation (4D),
    left foot site rotation (4D),
    right up leg rotation (4D),
    right leg rotation (4D),
    right foot rotation (4D),
    right foot site rotation (4D),
    tail rotation (4D),
    tail1 rotation (4D),
    tail1 site rotation (4D)
]

Positions are specified in meters and 3D rotations for spherical joints are specified as quaternions (w, x, y, z). The root positions and rotations are in world coordinates, but all other joint rotations are in the joint's local coordinates. To use your motion clip, convert it to a similar style JSON file.

Q: Crash with message "util/MathUtil.cpp:175: static tMatrix cMathUtil::RotateMat(const tVector &, double): Assertion std::abs(axis.squaredNorm() - 1) < 0.0001' failed." after throwing random boxes.

A: You can adjust the equation std::abs(axis.squaredNorm() - 1) < 0.0001 in file DeepMimic/DeepMimicCore/util/MathUtil.cpp, from value 0.0001 to 0.1, and re-compile CarlCore.

Q: Getting the error message ImportError: libBulletDynamics.so.2.88: cannot open shared object file: No such file or directory while running the codes.

A: Check libBulletDynamics.so.2.88 is presented under /usr/local/lib/ by running ls -l /usr/local/lib/ | grep libBullet, and then include it in the PATH by running export LD_LIBRARY_PATH=/usr/local/lib/.

We thank AI4Animation for providing the quadruped animation system and DeepMimic for providing the motion imitation framework.

This project is under the terms of the MIT license. It is only for research or education purposes, and not freely available for commercial use or redistribution. This intellectual property belongs to the Inventec Corporation. Licensing is possible if you want to use the code for commercial use. For scientific use, please reference this repository together with the relevant publication.

@article{CARL2020,
    author   = {Luo, Ying-Sheng and Soeseno, Jonathan Hans and Chen, Trista Pei-Chun and Chen, Wei-Chao},
    title    = {CARL: Controllable Agent with Reinforcement Learning for Quadruped Locomotion},
    journal  = {ACM Transactions on Graphics (Proceedings of SIGGRAPH 2020)},
    volume   = {39},
    number   = {4},
    year     = {2020},
    numpages = {10}
}