dreamerv3

Why we included this project

DreamerV3 is a JAX implementation of the DreamerV3 algorithm, a model-based reinforcement learning approach that has been written up in Nature. Instead of learning directly from trial and error, it builds a compressed world model from past experience and trains its policy on imagined future trajectories, which is why it can learn useful behaviors with comparatively little interaction. The headline result is that one fixed set of hyperparameters performs well across a wide range of tasks, from Atari and Minecraft to simulated robot control, so you can test an idea without tuning each environment separately. The repo ships with config files for the main benchmarks and a Docker setup, so getting a training run going and watching the metrics is straightforward. For anyone building control or embodied agents, it is a solid reference implementation to read, run, and compare against.