jdtcc/imitation

Code for the paper "Generative Adversarial Imitation Learning"

Status: Archive (code is provided as-is, no updates expected)

Generative Adversarial Imitation Learning

Jonathan Ho and Stefano Ermon

Contains an implementation of Trust Region Policy Optimization (Schulman et al., 2015).

Dependencies:

• OpenAI Gym >= 0.1.0, mujoco_py >= 0.4.0
• numpy >= 1.10.4, scipy >= 0.17.0, theano >= 0.8.2
• h5py, pytables, pandas, matplotlib

Provided files:

• expert_policies/* are the expert policies, trained by TRPO (scripts/run_rl_mj.py) on the true costs
• scripts/im_pipeline.py is the main training and evaluation pipeline. This script is responsible for sampling data from experts to generate training data, running the training code (scripts/imitate_mj.py), and evaluating the resulting policies.
• pipelines/* are the experiment specifications provided to scripts/im_pipeline.py
• results/* contain evaluation data for the learned policies