This is the code for the paper "Learning Constraints from Offline Demonstrations via Superior Distribution Correction Estimation" published at ICML 2024. The implementation is based on the code from Cleanrl and jax-rl.

You may run the following command to install dependencies:

pip install -r requirements.txt
pip install -U "jax[cuda]==0.4.23" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install git+https://github.com/quangr/dejax.git

Then you can download the dataset and put under the ./dataset.

to run Point Maze compare: jupyter nbconvert --to notebook --execute --inplace icrl/maze/exp/train.ipynb --ExecutePreprocessor.timeout=-1

To run Grid-World env with our method:

jupyter nbconvert --to notebook --execute --inplace icrl/grid_world/ICSDICE.ipynb --ExecutePreprocessor.timeout=-1

To run Grid-World env with RECOIL-V:

python icrl/grid_world/recoil-V.py

To run mujoco env with our method:

python icrl/benchmark/ICSDICE.py --seed 1 --alpha=0.0001 --total-timesteps=1000000 --debug=False --update-period=100000 --beta=0.5 --env-id=Ant_ls --cost-l2=0.0005 --cost-limit=0.9 --expert-ratio 1.0