This is the official PyTorch implementation of SMODICE: Versatile Offline Imitation Learning via State Occupancy Matching, presented at International Conference on Machine Learning, 2022.

1. Offline Imitation Learning from Mismatched Experts

python smodice_tabular/run_tabular_mismatched.py

2. Offline Imitation Learning from Examples

python smodice_tabular/run_tabular_example.py

1. Create conda environment and activate it:

   conda env create -f environment.yml
   conda activate smodice
   pip install --upgrade numpy
   pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
   git clone https://github.com/rail-berkeley/d4rl
   cd d4rl
   pip install -e .
   
   

1. Run the following command with variable ENV set to any of hopper, walker2d, halfcheetah, ant, kitchen.

python run_oil_observations.py --env_name $ENV

2. For the AntMaze environment, first generate the random dataset:

cd envs
python generate_antmaze_random.py --noise

Then, run

python run_oil_antmaze.py

1. For halfcheetah and ant, run

python run_oil_observations.py --env_name halfcheetah --dataset 0.5 --mismatch True

and

python run_oil_observations.py --env_name ant --dataset disabled --mismatch True

respectively. 2. For AntMaze, run

python run_oil_antmaze.py --mismatch True

1. For the PointMass-4Direction task, run

python run_oil_examples_pointmass.py

2. For the AntMaze task, run

python run_oil_antmaze.py --mismatch False --example True

3. For the Franka Kitchen based tasks, run

python run_oil_examples_kitchen.py --dataset $DATASET

where DATASET can be one of microwave, kettle.

For any task, the BC baseline can be run by appending --disc_type bc to the above commands.

For RCE-TD3-BC and ORIL baselines, on the appropriate tasks, append --algo_type $ALGO where ALGO can be one of rce, oril.

If you find this repository useful for your research, please cite

@article{ma2022smodice,
      title={SMODICE: Versatile Offline Imitation Learning via State Occupancy Matching}, 
      author={Yecheng Jason Ma and Andrew Shen and Dinesh Jayaraman and Osbert Bastani},
      year={2022},
      url={https://arxiv.org/abs/2202.02433}
}

If you have any questions regarding the code or paper, feel free to contact me at jasonyma@seas.upenn.edu.

This codebase is partially adapted from optidice, rce, relay-policy-learning, and d4rl ; We thank the authors and contributors for open-sourcing their code.