Object-Centric Scene Representations using Active Inference

This repository contains the code for the experiments used in Object-Centric Scene Representations using Active Inference by Toon Van de Maele, Tim Verbelen, Pietro Mazzaglia, Stefano Ferraro, and Bart Dhoedt.

Code structure

• scene-environments: contains the code for the environments
• The code for the model and the active inference agent is in ccn
• dommeltk is a framework we built around pytorch to aid us in development and is required to run the ccn models.

Download data assets & trained models

The trained models/datasets/evaluation scenes can be downloaded in the following link.

To run the benchmark, please move the downloaded files into the following directories:

- ccn-data -> ccn/data
- scene-environments-data -> scene-environments/data 

Citation

If you find the code useful, please refer to our work using:

@misc{vandemaele2023objectcentricscenerepresentationsusing,
      title={Object-Centric Scene Representations using Active Inference}, 
      author={Toon Van de Maele and Tim Verbelen and Pietro Mazzaglia and Stefano Ferraro and Bart Dhoedt},
      year={2023},
      eprint={2302.03288},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2302.03288}, 
}

Acknowledgements

The dommel toolkit was co-developed at Ghent University by Tim Verbelen @tverbele, Ozan Catal @ozancatal, and Toon Van de Maele @toonvdm.