dylandjian/retro-contest-sonic

This project is a historical artifact from the 2018 OpenAI Retro Contest. While it implements the influential 'World Models' paper by Ha and Schmidhuber, it is highly specific to the Sonic the Hedgehog environment and uses a deprecated stack (likely TensorFlow 1.x). With only 78 stars and zero velocity over several years, it serves as a personal reference rather than a living tool. The defensibility is near-zero as the methodology has been superseded by more advanced model-based RL architectures like Dreamer (V1-V3) and IRIS, which offer better sample efficiency and generalizability. Frontier labs have no interest in this specific niche, as their focus has shifted from specialized game-solvers to generalist agents. It lacks a moat because the code is a standard application of known research to a static benchmark that is no longer the industry standard for RL evaluation.