World Model

PyTorch implementation of world model.

VAE - encodes the input data (images) into a latent distribution that captures the features of the data (tracks, car, etc. in this case).

MDN-RNN - LSTM with a Mixture Density Network head that predicts the next latent state given the current state and action.

Controller - A single linear layer mapping (z_t, h_t) to 3 continuous actions: steering (tanh), gas (sigmoid), brake (sigmoid).

Project Structure

world-model/
  modules/          # Model definitions (VAE, MDN-RNN, Controller)
  dataset/          # Data collection and dataset classes
  train/            # Training scripts

Citation

@article{https://doi.org/10.5281/zenodo.1207631,
  doi = {10.5281/ZENODO.1207631},
  url = {https://zenodo.org/record/1207631},
  author = {Ha, David and Schmidhuber, Jürgen},
  title = {World Models},
  publisher = {Zenodo},
  year = {2018},
  copyright = {Creative Commons Attribution 4.0}
}