This class is based on ppo1 from the openai baselines but removes a bit of bloat and a few unnecessarily complicated algorithmic designs.
The model is an MLP with two hidden layers of 64 nodes so it doesn't take much computing to reach a respectable level of performance. I ran the training algorithm for about 10 mins in a single process after which the half cheetah achieved a decent looking gait.
The main two dependencies are Roboschool and OpenAI baselines for which the installation instructions are here and here respectively.
python main.py --env=$YOUR_ENV_NAME --ntimesteps=$PROBABLY_ABOUT_TWENTY_MILL
python main.py --env=$YOUR_ENV_NAME --ntimesteps=$PROBABLY_ABOUT_TWO_THOU --train=False