About
Dreamer is a family of model-based reinforcement learning (RL) algorithms developed by Danijar Hafner and collaborators at Google Brain, Google DeepMind and the University of Toronto. The agent learns a representation of the environment from observations (such as images) using a Recurrent State-Space Model (RSSM) and trains an actor-critic policy on trajectories rolled out in imagination — without executing actions in the real environment. Successive versions include PlaNet (2019, model-based planning), Dreamer (2019), DreamerV2 (2020, discrete latent representations, strong Atari 2600 results) and DreamerV3 (2023/2025, a single set of hyperparameters for more than 150 tasks and the first agent to collect diamonds in Minecraft without human data). The whole family is released as open source under the MIT license and represents one of the key research directions in world models and general-purpose RL algorithms.