danijar/dreamerv3

Builds a latent world model using categorical representations that predicts future states and rewards, then trains an actor-critic policy through imagined rollouts. Implemented in JAX with support for diverse environments (Atari, robotics, visual control tasks) and scales efficiently—larger models improve both final performance and sample efficiency without hyperparameter tuning across domains.

2,917 stars and 412 monthly downloads. No commits in the last 6 months. Available on PyPI.