taapasx28 / dm-control-starter-kit Goto Github PK

Most of the installation procedure for dm_control is listed in the official repository here. However some key points to take care of are:

1. Extract Mujoco pro in ~/.mujoco/ directory. The key mjkey.txt you with get for Mujoco Pro 2.0 needs to be placed in the bin directory to run Mujoco independently. A sanity check that Mujoco is installed correctly can be done by ./simulate ../model/humanoid.xml command (Linux and MacOS) by going into the bin directory. However dm_control requires the key in the .mujoco directory so place it there too.

2. dm_control can then be installed by pip as stated in the repository or by cloning the repository and installing that by pip intall -e.

3. For rendering although the dependencies listed for linux are libglfw3 libglew2.0 these might not work in that case use sudo apt-get install libglfw-dev libglew-dev. These substitute dependencies work out.

4. To test out dm_control try running the code testdm.py. If everything is installed correctly it will print out the reward, discount and observation of an episode for a simple cartpole task. Unlike gym after a step these three things are returned as a timestep object form which they can be accessed.

For training tasks in dm_control I recommend the tonic RL library that can be found here. The installation is pretty straight forward. The library provides with 2 variants of RL algorithms in Pytorch as well as tensorflow so it can be modified easily for research. Regarding how to train go to training folder of the repository.

• For how to train a control suite environment go to .\train\ and for making a custom environment refer to .\create_env\