[ 📺 Website | 🏗 Github Repo | 🎓 Paper ]

• Gym >= 0.8.1

• Mujoco-py >= 0.5.7

• Tensorflow >= 1.0.1

• Mujoco

  Add the following path to ~/.bashrc

  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco200/bin
  export LD_LIBRARY_PATH=$HOME/.mujoco/mjpro150/bin:$LD_LIBRARY_PATH
  

  Follow instructions to install mujoco_py v1.5 here.

• SenseAct (Optional)

  SenseAct uses Python3 (>=3.5), and all other requirements are automatically installed via pip.

  On Linux and Mac OS X, run the following:

  git clone https://github.com/kindredresearch/SenseAct.git
  cd SenseAct
  pip install -e .
  

1. Collect demonstration data and save to expert_data directory.

The expert data should be a python pickle file (with .bin but not .pkl as a suffix) It has batch_size, action, states (required by set_er_stats()), like the expert_data/hopper_er.bin (just as an example).

2. Training

COUNTER=1
ENVS+='Ant-v2'
for ENV_ID in ${ENVS[@]}
do
  CUDA_VISIBLE_DEVICES=`expr $COUNTER % 4` python main.py --env_name $ENV_ID --alg mairlImit --obs_mode state &
  COUNTER=$((COUNTER+1))
done

3. Evaluation

CUDA_VISIBLE_DEVICES=0 python main.py --env_name Ant-v2 --train_mode False

Our code is based on itaicaspi/mgail, HumanCompatibleAI/imitation, huggingface/transformers.

Adversarial Inverse Reinforcement Learning with Self-attention Dynamics Model
Jiankai Sun, Lantao Yu, Pinqian Dong, Bo Lu, and Bolei Zhou
In IEEE Robotics and Automation Letters (RA-L) 2021
[Paper] [Project Page]

@ARTICLE{sun2021adversarial,
     author={J. {Sun} and L. {Yu} and P. {Dong} and B. {L} and B. {Zhou}},
     journal={IEEE Robotics and Automation Letters},
     title={Adversarial Inverse Reinforcement Learning with Self-attention Dynamics Model},
     year={2021},
}