Intro

Code accompanying the SIGGRAPH 2018 paper: "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills". The framework uses reinforcement learning to train a simulated humanoid to imitate a variety of motion skills from mocap data.

Project page: https://xbpeng.github.io/projects/DeepMimic/index.html

Dependencies

C++:

Bullet 2.87 (https://github.com/bulletphysics/bullet3/releases)
Eigen (http://www.eigen.tuxfamily.org/index.php?title=Main_Page)
OpenGL >= 3.2
freeglut (http://freeglut.sourceforge.net/)
glew (http://glew.sourceforge.net/)

Python:

Python 3
PyOpenGL (http://pyopengl.sourceforge.net/)
Tensorflow (https://www.tensorflow.org/)
MPI4Py (https://mpi4py.readthedocs.io/en/stable/install.html)

Misc:

SWIG (http://www.swig.org/)
MPI
- Windows: https://docs.microsoft.com/en-us/message-passing-interface/microsoft-mpi
- Linux: sudo apt install libopenmpi-dev

Build

The simulated environments are written in C++, and the python wrapper is built using SWIG. To install the python dependencies, run

pip install -r requirements.txt

Note that MPI must be installed before MPI4Py. When building Bullet, be sure to disable double precision with the build flag USE_DOUBLE_PRECISION=OFF.

Windows

The wrapper is built using DeepMimicCore.sln.

Select the x64 configuration from the configuration manager.
Under the project properties for DeepMimicCore modify Additional Include Directories to specify
- Bullet source directory
- Eigen include directory
- python include directory
Modify Additional Library Directories to specify
- Bullet lib directory
- python lib directory
Build DeepMimicCore project with the Release_Swig configuration and this should generate DeepMimicCore.py in DeepMimicCore/.

Linux

Modify the Makefile in DeepMimicCore/ 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
Build wrapper,
```
make python
```

This should generate DeepMimicCore.py in DeepMimicCore/

How to Use

Once the python wrapper has been built, training is done entirely in python using Tensorflow. DeepMimic.py runs the visualizer used to view the simulation. Training is done with mpi_run.py, which uses MPI to parallelize training across multiple processes.

DeepMimic.py is run by specifying an argument file that provides the configurations for a scene. For example,

python DeepMimic.py --arg_file args/run_humanoid3d_spinkick_args.txt

will run a pre-trained policy for a spinkick. Similarly,

python DeepMimic.py --arg_file args/kin_char_args.txt

will load and play a mocap clip.

To train a policy, run mpi_run.py by specifying an argument file and the number of worker processes. For example,

python mpi_run.py --arg_file args/train_humanoid3d_spinkick_args.txt --num_workers 4

will train a policy to perform a spinkick using 4 workers. As training progresses, it will regularly print out statistics and log them to output/ along with a .ckpt of the latest policy. It typically takes about 60 millions samples to train one policy, which can take a day when training with 16 workers. 16 workers is likely the max number of workers that the framework can support, and it can get overwhelmed if too many workers are used.

A number of argument files are already provided in args/ for the different skills. train_[something]_args.txt files are setup for mpi_run.py to train a policy, and run_[something]_args.txt files are setup for DeepMimic.py to run one of the pretrained policies. To run your own policies, take one of the run_[something]_args.txt files and specify the policy you want to run with --model_file. Make sure that the reference motion --motion_file corresponds to the motion that your policy was trained for, otherwise the policy will not run properly.

Interface

the plot on the top-right shows the predictions of the value function
right click and drag will pan the camera
left click and drag will apply a force on the character at a particular location
scrollwheel will zoom in/out
pressing 'r' will reset the episode
pressing 'l' will reload the argument file and rebuild everything
pressing 'x' will pelt the character with random boxes
pressing space will pause/resume the simulation
pressing '>' will step the simulation one step at a time

Mocap Data

Mocap clips are located in data/motions/. To play a clip, first modify args/kin_char_args.txt and specify the file to play with --motion_file, then run

python DeepMimic.py --arg_file args/kin_char_args.txt

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),
	chest rotation (4D),
	neck rotation (4D),
	right hip rotation (4D),
	right knee rotation (1D),
	right ankle rotation (4D),
	right shoulder rotation (4D),
	right elbow rotation (1D),
	left hip rotation (4D),
	left knee rotation (1D),
	left ankle rotation (4D),
	left shoulder rotation (4D),
	left elbow rotation (1D)
]

Positions are specified in meters, 3D rotations for spherical joints are specified as quaternions (w, x, y ,z), and 1D rotations for revolute joints (e.g. knees and elbows) are represented with a scalar rotation in radians. The root positions and rotations are in world coordinates, but all other joint rotations are in the joint's local coordinates. To use your own motion clip, convert it to a similar style JSON file.

colijian/DeepMimic