eloialonso/iris

Built on a two-stage architecture combining a VQ-VAE discrete autoencoder with an autoregressive Transformer, IRIS reformulates world modeling as sequence prediction over learned image tokens rather than raw pixels. Trained end-to-end with a reinforcement learning actor-critic on Atari environments, it generates millions of imagined rollouts for policy optimization while requiring minimal real environment interaction. The codebase integrates with PyTorch, Hydra for configuration management, and Weights & Biases for experiment tracking, with pretrained checkpoints available on Hugging Face.

870 stars. No commits in the last 6 months.