https://github.com/ubisoft/ubisoft-laforge-sciql
Codebase associated with the paper Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment
https://github.com/ubisoft/ubisoft-laforge-sciql
Last synced: 6 days ago
JSON representation
Codebase associated with the paper Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment
- Host: GitHub
- URL: https://github.com/ubisoft/ubisoft-laforge-sciql
- Owner: ubisoft
- License: other
- Created: 2026-06-29T13:05:07.000Z (26 days ago)
- Default Branch: master
- Last Pushed: 2026-07-06T02:08:31.000Z (19 days ago)
- Last Synced: 2026-07-19T00:36:54.497Z (6 days ago)
- Language: Python
- Size: 118 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE.txt
Awesome Lists containing this project
README
## Overview
This repository is the official implementation of the **Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment** paper published at ICML 2026 (Spotlight).
## Installation
1) Create and activate conda environment:
```
conda create -n sciql python=3.10 -y && conda activate sciql
```
2) Install pip dependencies:
```
python -m pip install --no-cache-dir -r requirements.txt
```
3) Setup the environment variables:
```
conda env config vars set PYTHONPATH="$PYTHONPATH:$PWD"
conda env config vars set MUJOCO_GL=egl
conda deactivate && conda activate sciql
```
4) Download the datasets from our website. Organize them in a datasets/ folder at the root of the project:
```
datasets/
└── diverse_mujoco/
└── mujoco_halfcheetah-fix/
└── fix-val.npz
└── fix.npz
└── mujoco_halfcheetah-stitch/
└── stitch-val.npz
└── stitch.npz
└── mujoco_halfcheetah-vary/
└── vary-val.npz
└── vary.npz
└── traj2d
└── random_circles-inplace-v0/
└── random_circles-navigate-v0/
```
## Usage
To reproduce the results, launch the experiments with the following commands:
```
# For Halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/bc/jax -cn mujoco_halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/cbc/jax -cn mujoco_halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/bcpmi_joint/jax -cn mujoco_halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/scbc/jax -cn mujoco_halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/sciql_joint/jax -cn mujoco_halfcheetah
python experiments/control/launch.py -cp yamls/diverse_mujoco/sorl/jax -cn mujoco_halfcheetah
# For Circle2D
python experiments/control/launch.py -cp yamls/traj2d/bc/jax -cn random_circles-v0
python experiments/control/launch.py -cp yamls/traj2d/cbc/jax -cn random_circles-v0
python experiments/control/launch.py -cp yamls/traj2d/bcpmi_joint/jax -cn random_circles-v0
python experiments/control/launch.py -cp yamls/traj2d/scbc/jax -cn random_circles-v0
python experiments/control/launch.py -cp yamls/traj2d/sciql_joint/jax -cn random_circles-v0
python experiments/control/launch.py -cp yamls/traj2d/sorl/jax -cn random_circles-v0
```
## Aknownledgments
This codebase takes inspiration from the [jax_corl](https://github.com/nissymori/JAX-CORL) library.
## Citation
Accepted at ICML 2026. Proceedings citation will be updated once available.
```
@inproceedings{petitbois2026offline,
title = {Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment},
author = {Petitbois, Mathieu and Portelas, R{\'e}my and Lamprier, Sylvain},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2026}
}
```
© [2026] Ubisoft Entertainment. All Rights Reserved.