xiaoyuanzh/BGAIL

Bayesian Approach to Generative Adversarial Imitation Learning

Implementation of Bayesian Generative Adversarial Imitation Learning (BGAIL)

Requirements

• python 3.6.6
• MuJoCo 1.31
• mujoco-py 0.5.7
• OpenAI Gym 0.9.0
• OpenAI Baselines 0.1.5 (in this repository)
• TensorFlow 1.10.0
• See here for detailed procedure.

References

• openai/imitation
  • GAIL implementation of authors
• andrewliao11/gail-tf
  • TensorFlow implementation of GAIL
• OpenAI baselines
  • Baseline implementation of GAIL
• wsjeon/SVGD
  • TensorFlow implementation of Stein variational gradient descent (SVGD)
• BGAIL paper

Usage

Download expert trajectories.

We use the expert trajectories by using the code given by openai/imitation.

1. Download expert trajectories from this link to expert_trajs/.
2. Run

   python expert_trajs/convert_h5_to_pkl.py

   to convert expert trajectories into *.pkl files.

Run.

python run.py 

will run BGAIL in default setting (see bgail/defaults.py.)