Genie Redux

Introduction

GenieRedux is the official model repository introduced in the paper "Learning Generative Interactive Environments By Trained Agent Exploration" presented at the NeurIPS'24 D3S3 Workshop. It is an open implementation of the Genie world model, with GenieRedux-G as its variant adapted for virtual environments.

Architecture

The GenieRedux model includes:

• A spatial-temporal tokenizer with a VQ codebook.
• GenieRedux: Combines Dynamics and LAM (Latent Action Model), utilizing action embedding from LAM.
• GenieRedux-G: Focuses on dynamics alone, employing one-hot action input.

Training

GenieRedux models were trained using the CoinRun environment, as per the original study's case framework. The tokenizer and models leverage dynamics and action embeddings to explore and learn within generative interactive environments.

Guide: Running Locally

1. Clone the Repository:
   Clone the official code repository from GitHub.

2. Install Dependencies:
   Ensure all required dependencies and packages are installed as per the provided requirements.txt or setup instructions.

3. Run the Model:
   Follow any additional instructions in the repository for executing the model locally.

• Cloud GPUs: For enhanced performance, consider using cloud GPU services such as AWS, Google Cloud, or Azure.

License

The GenieRedux model is licensed under the MIT License.