RIZE is a novel IRL method with adaptive reward bounds and distributional critics, achieving expert-level performance on MuJoCo and Adroit tasks using only 3–10 expert demonstrations. It outperforms IQ-Learn, LSIQ, SQIL, CSIL, and BC—especially on high-DoF tasks like Humanoid-v2 and Hammer-v1. See the project page for results, and visuals.
-
Clone the repository
git clone https://github.com/adibka/RIZE.git cd RIZE -
Create an environment (Python 3.10)
python3 -m venv venvs/rize source venvs/rize/bin/activate -
Install mujoco-py
Note:
mujoco-py==2.0.2.0requires MuJoCo version 2.0 binaries. Follow the guide at mujoco-py docs if needed.mkdir -p ~/.mujoco && cd ~/.mujoco curl -O https://www.roboti.us/download/mujoco200_linux.zip curl -O https://www.roboti.us/file/mjkey.txt unzip mujoco200_linux.zip mv mujoco200_linux mujoco200 rm mujoco200_linux.zip
Add to
~/.bashrc(replace{user-name}with your actual username):echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/{user-name}/.mujoco/mujoco200/bin' >> ~/.bashrc source ~/.bashrc
-
Install dependencies
pip install -r requirements.txt
Note:
requirements.txtpins tested versions (includingmujoco-py,gym,gymnasium,mujoco). Verified on Ubuntu 22.04 + Python 3.10.
Download pre-generated expert trajectories:
cd RIZE
./download_experts.shRun training with a single command. Hyperparameters are detailed in the paper appendix (Table 1).
python main.py --env halfcheetah --demos 10 --seed 0Replace
--envwith any supported task:
halfcheetah,walker2d,ant,humanoid,hopper,hammer
@article{karimi2025rize,
title={RIZE: Adaptive Regularization for Imitation Learning},
author={Karimi, Adib and Ebadzadeh, Mohammad Mehdi},
journal={Transactions on Machine Learning Research},
year={2025},
url={https://openreview.net/forum?id=a6DWqXJZCZ}
}