Offline Learning of Controllable Diverse Behaviors

Paper

Overview

This repository is the official implementation of the Offline Learning of Controllable Diverse Behaviors paper published at the Generative Models for Robot Learning Workshop at ICLR 2025.

Installation

1. Create and activate conda environment:

conda create -n swr python=3.10 -y -q && conda activate swr

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 deactivate && conda activate swr

4. Download the datasets from: [link]. Organize them in a datasets/ folder at the root of the project:

datasets/
└── d3il/
    └── data/
        └── aligning/
            └── split_data.py
            └── train_files.pkl
            └── eval_files.pkl
            └── all_data/
            └── ...
        └── avoiding
            └── data/
└── maze
    └── medium_maze-only_forward-human/
    └── medium_maze-one_side-human/

Usage

To reproduce the results, launch the experiments with the following commands:

# Launch the training
bash swr/scripts/local/training/d3il_avoiding/swr_wzbc_benchmark.sh

# Launch the diversity experiment
bash swr/scripts/local/experiments/diversity/d3il_avoiding/swr_wzbc_exp.sh

# Launch the control experiment
bash swr/scripts/local/experiments/control/d3il_avoiding/swr_wzbc_exp.sh

Acknowledgments

This codebase uses code from the d3il library.

Citation

@inproceedings{petitbois2025offlinelearningcontrollablediverse,
      title={Offline Learning of Controllable Diverse Behaviors}, 
      author={Mathieu Petitbois and R{\'e}my Portelas and Sylvain Lamprier and Ludovic Denoyer},
      booktitle={ICLR 2025 Workshop on Generative Models for Robot Learning},
      year={2025},
      note={Workshop paper},
      url={https://arxiv.org/abs/2504.18160}
}

© [2026] Ubisoft Entertainment. All Rights Reserved.