Mastering Diverse Domains through World Models

ABSTRACT

General intelligence requires solving tasks across diverse domains without human intervention. We present DreamerV3, a general and scalable reinforcement learning algorithm that masters a wide range of domains with fixed hyperparameters. DreamerV3 learns a world model from environmental interactions and trains an actor-critic policy entirely from imagined trajectories predicted by the world model. It outperforms previous approaches across 150 tasks spanning continuous control, visual navigation, and discrete Atari games. Notably, DreamerV3 is the first algorithm to collect diamonds in Minecraft entirely from scratch without using human demonstrations or hand-crafted curricula.