dreamerv2

Why we included this project

DreamerV2 is the agent that showed world-model methods could hit human-level scores on the Atari benchmark, and this repository is its reference implementation in TensorFlow 2. Instead of learning a policy directly from reward signals, the agent builds a compact latent model of the environment from raw image observations, then trains actor and critic networks on rollouts imagined through that model. That approach buys sample efficiency, and the included training curves across all 55 Atari games let you verify it yourself rather than taking the paper's word for it. The package installs with pip and the train and evaluation loop runs on a single GPU, so reproducing results is realistic on modest hardware, and a Dockerfile takes care of dependency setup. For anyone digging into world models or sample-efficient control, reading this codebase will teach you more than most tutorials.