tgangwani / BMIL

Pytorch code for "Learning Belief Representations for Imitation Learning in POMDPs" (UAI 2019)

This repo contains code for our paper Learning Belief Representations for Imitation Learning in POMDPs published at UAI 2019.

The code was tested with the following packages:

• python 3.6.6
• pytorch 0.4.1
• gym 0.10.8

Running command

To run MuJoCo experiments, use the script run_mujoco.sh with the following usage:

bash run_mujoco.sh [env] [belief_loss_type] [belief_regularization]

BMIL results can be reproduced with bash run_mujoco.sh [env] task_aware True

Expert trajectories

Please update the path to expert trajectories in the file "code/conf/envParams.yaml". Also see the storage requirements in "code/expert_envs.py" and modify as per convenience.

Credits

The base for this code is provided by DVRL, which itself utilizes methods from this. We also use OpenAI baselines helpers for vectorized environments.