DreamerV3 in PyTorch, but the maintainers moved on

What it does This is a PyTorch port of DreamerV3, the world-model-based reinforcement-learning algorithm that learns to plan across vision, proprioception, and memory tasks. The repo supports benchmarks including DeepMind Control, Atari 100k, Crafter, Minecraft, and Memory Maze, and includes training and evaluation scripts. The authors note it was implemented prior to major DreamerV3 updates and does not reflect them.

The interesting bit The maintainers are refreshingly blunt: they flag this codebase as outdated and point to their new r2dreamer repository as a roughly five-times-faster, up-to-date alternative. That makes this repo less a destination and more a snapshot of an earlier PyTorch port.

Key highlights

• PyTorch reimplementation of the DreamerV3 world-model RL agent
• Covers continuous and discrete control across vision, proprioception, and memory tasks
• Includes benchmark result plots for DMC, Atari 100k, and Crafter
• Authors now direct users to r2dreamer for the latest DreamerV3 baseline

Caveats

• Explicitly marked as outdated; predates major DreamerV3 updates, which the README says accounts for several GitHub Issues
• OpenGL headless rendering issues reported when running DMC tasks
• Heavily inspired by existing JAX and TensorFlow implementations

Verdict Researchers tracing the evolution of PyTorch Dreamer ports might find this useful; everyone else should follow the authors’ own advice and use r2dreamer instead.