Object-Oriented Dynamics Predictor (OODP)

This project is a re-implementation of the Object-Oriented Dynamics Predictor (OODP) originally developed by Zhu et al. (Guangxiang Zhu*, Jianhao Wang*, Zhizhou Ren*, and Chongjie Zhang, "Object-Oriented Dynamics Learning through Multi-Level Abstraction", Deep RL Workshop NIPS 2018.) which used Tensorflow 1 Github project of Zhu et al.. The goal of OODP is to predict the future dynamics of objects in a scene using deep learning techniques.

This implementation of OODP is based on TensorFlow 2 and was completed as part of the Neural Network Seminar at the University of Hamburg during the Summer semester of 2020.

Warning: The Game Monsterkong is currently not working because of the missing PygameWrapper lib.

Dependencies

The project requires the following dependencies:

• TensorFlow
• NumPy
• tqdm
• logging
• Pillow
• PyGame

These dependencies can be installed using Pipenv. Please refer to the Pipfile for specific versions.

Usage

1. Set up a Python environment with the required dependencies (e.g., using Pipenv).
2. Download and init Monsterkong from the PyGame-Learning-Environment.
3. Create datasets using the createDataset.py script.
4. Train the models using the train.py or train_dist.py scripts.
5. Use the trained models for prediction and analysis using the analysis.ipynb notebook.
6. Refer to the documentation in the docs/ directory for additional information and resources.

Results

The results are documented here.

Acknowledgments

We would like to express our gratitude to Zhu et al. for developing the original Object-Oriented Dynamics Predictor (OODP) and providing the inspiration for this re-implementation. We would also like to thank the Neural Network Seminar at the University of Hamburg for the opportunity to work on this project.