Intro
Source code for the paper: Terrain-Adaptive Locomotion Skills using Deep Reinforcement Learning https://www.cs.ubc.ca/~van/papers/2016-TOG-deepRL/index.html
Setup
This section covers some of the steps to setup and compile the code. The software depends on many libraries that need to be carefully prepared and placed for the building and linking to work properly.
Linux
Dependencies
-
Caffe (http://caffe.berkeleyvision.org/installation.html)
Specific version (https://github.com/niuzhiheng/caffe.git @ 7b3e6f2341fe7374243ee0126f5cad1fa1e44e14) sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
sudo apt-get install libatlas-base-devIn the instruction to make and build Caffe uncomment the CPU only line
# CPU-only switch (uncomment to build without GPU support). CPU_ONLY := 1
Or if on Windows
https://github.com/initialneil/caffe-vs2013 -
Boost
-
OpenCV
-
BulletPhysics
-
CUDA
Package Manager Installation
Install repository meta-data
When using a proxy server with aptitude, ensure that wget is set up to use the same proxy settings before installing the cuda-repo package.
$ sudo dpkg -i cuda-repo-.deb
Update the Apt repository cache
$ sudo apt-get update
Install CUDA
$ sudo apt-get install cuda -
Eigen (http://eigen.tuxfamily.org/index.php?title=Main_Page)
-
bits
sudo apt-get install gcc-4.9-multilib g++-4.9-multilib
Linux Build Instructions
- Download the most recent compressed external file from the newest release.
- Extract it and move into the DeepTerrainRL directory.
- Rebuild caffe
- Cd into external/caffe
- Make clean
- Make
- Cd back to ../../
- Copy the caffe libraries from external/caffe/build/lib to ./lib
- Premake4 clean
- Premake4 gmake
- Make config=debug64
- Everything should build fine.
Note: There are some issues with the installation on Ubuntu 16.04. Some of the libraries have changed their location and name (see BVLC/caffe#2347 for a solution).
Windows
This setup has been tested on Windows 7 and 10 with visual studio 2013.
- Download the library.zip file that contains almost all of the relevant pre compiled external libraries and source code.
- Unpack this library in the same directory the project is located in. For example, TerrainRL/../.
- You might need to install opengl/glu/GL headers. We have been using freeglut for this project. glew might already be included in library.zip.
- You will need to copy some dll files from dynamic_lib.zip to the directory the project is compiled to. For example, optimizer/x64/Debug/. These files are needed by the framework during runtime.
- Might need to create a folder in TerrainRL called "output", This is where temprary and current policies will be dumped.
Runing The System
After the system has been build there are two executable files that server different purposes. The TerrainRL program is for visually simulating the a controller and TerrainRL_Optimize is for optimizing the parameters of some controller.
Examples:
To simulate a controller/character
./TerrainRL -arg_file= args/sim_dog_args.txt
To simulate a controller/character with a specific policy
./TerrainRL_Optimizer -arg_file= args/dog_slopes_mixed_args.txt
To Train a controller
./TerrainRL_Optimizer -arg_file= args/opt_args_train_mace.txt
Key Bindings
Most of these are togglesg
-
c fixed camera mode
-
y draw COM path and contact locations
-
q draw "filmstrip" like rendering
-
f draw torques
-
h draw Actor value functions and feature visualization
-
shift + '>' step one frame
-
p toggle draw value function
-
',' and '.' change render speed, decrease and increase.
-
"spacebar" to pause simulation
-
r restart the scenario
-
l reload the simulation (reparses the arg file)
-
g draw state features
-
x spawn projectile
-
z spawn big projectile
-
click on character and drag to apply force