Dreamer (learning by latent imagination)

Dreamer is the idea, taken to its logical end, that an agent can learn how to act almost entirely inside its own head. It first learns a latent world model from real experience, and then trains its policy not on the real environment at all but on long sequences of futures imagined in that latent space — it learns to act well in the dream, and that skill transfers to reality.

The world model encodes observations into latent states and predicts their evolution, rewards, and values. To improve behaviour, Dreamer rolls the latent model forward for many steps from sampled starting states, producing imagined trajectories, and backpropagates the predicted long-term return through the differentiable dynamics straight into the policy — an actor-critic trained on dreams. Real interaction is used only to refine the world model, making it strikingly data-efficient on tasks from pixels.

Dreamer is a landmark because it showed that learning purely from imagination can match or beat model-free agents on hard continuous-control and Atari tasks while using far less real data. Its limits are the usual ones: the dream is only as trustworthy as the world model that generates it.