i-m-iron-man / libtorch_mujoco_DDPG

DDPG algorithm on an inverted pendulum environment using Mujoco(C++) and pytorch C++ front end.

Inverted Pendulum task

The task is to balance the pendulum upright using continuous control when the maximum torque available is not enough to overcome gravity. In this implementation a DDPG algorithm is used to learn the task. Mujoco libarary(http://www.mujoco.org/) is used to model the environment and pytorch c++ frontend(https://pytorch.org/cppdocs/frontend.html) provides the support for reinforcement learning framework.

Platform

• OS: windows10(64 bit)
• CUDA: 10.1
• cuDNN: 7.6.3
• Compiler: MSVC 19.28.29336.0
• C++17
• Cmake: 3.19.2

Dependencies

• Mujoco: mujoco200 win64
• pytorch C++ frontend: 1.7.1 release
• boost: 1.75.0

Instructions

• In CMakeLists.txt replace C:/Users/kings/Desktop/source/mujoco200_win64/mujoco200_win64 at line 13 with the path to your mujoco installation
• In CMakeLists.txt replace C:/Users/kings/Desktop/source/libtorch-mujoco-git at line 14 with the path to where you download these source files
• open a command-prompt and cd to the location of these files
• mkdir build && cd build
• In order to build use command cmake -DCMAKE_PREFIX_PATH=<absolute path to libtorch>\libtorch ..
• In order to make use command cmake --build . --config Release
• In order to run first cd Release then libtorch-mujoco-git.exe ..\..\model\simple_pendulum.xml

Inspiration

https://github.com/EmmiOcean/DDPG_LibTorch