iris

Why we included this project

IRIS is the codebase behind an ICLR 2023 paper that made a case for learning world models with far less experience than earlier approaches demanded. Its agent learns to play Atari games from about two hours of real gameplay, a fraction of what classic RL agents need. The key idea is a discrete autoencoder that turns game frames into token sequences, paired with an autoregressive transformer that predicts future states and lets the agent rehearse millions of imagined trajectories. The codebase uses Hydra for configuration and supports checkpointing and Weights & Biases logging, which makes it a practical reference for reproducing the paper or extending the approach. Anyone working on model-based RL or world models will get up to speed faster here than from the paper alone.