Junewind / Code-for-Error-Bounds-of-Imitating-Policies-and-Environments

Error Bounds of Imitating Policies and Environments

This is the repository hosting the code used for our NeurIPS 2020 paper: Error Bounds of Imitating Policies and Environments. The code contains the implementation of the BC, GAIL, DAgger, FEM, MWAL, MBRL_BC, MBRL_GAIL.

Requirements

We use Python 3.6 to run all experiments. Please install MuJoCo following the instructions from mujoco-py. Other python packages are listed in requirement.txt

Dataset

Dataset, including expert demonstrations and expert policies (parameters), is provided in the folder of dataset.

However, one can run SAC to re-train expert policies (see scripts/run_sac.sh) and to collect expert demonstrations (see scripts/run_collect.sh).

Usage

The folder of scripts provides all demo running scripts to test algorithms like GAIL, BC, DAgger, FEM, GTAL, and imitating-environments algorithms.